diff --git a/scripts/examples/25-Machine-Learning/nn_cifar10_search_just_center.py b/scripts/examples/25-Machine-Learning/nn_cifar10_search_just_center.py
new file mode 100644
index 000000000..8afa2a06d
--- /dev/null
+++ b/scripts/examples/25-Machine-Learning/nn_cifar10_search_just_center.py
@@ -0,0 +1,53 @@
+# CIFAR-10 Search Just Center Example
+#
+# CIFAR is a convolutional nueral network designed to classify it's field of view into several
+# different object types and works on RGB video data.
+#
+# In this example we slide the LeNet detector window over the image and get a list of activations
+# where there might be an object. Note that use a CNN with a sliding window is extremely compute
+# expensive so for an exhaustive search do not expect the CNN to be real-time.
+
+import sensor, image, time, os, nn
+
+sensor.reset() # Reset and initialize the sensor.
+sensor.set_pixformat(sensor.RGB565) # Set pixel format to RGB565 (or GRAYSCALE)
+sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
+sensor.set_windowing((128, 128)) # Set 128x128 window.
+sensor.skip_frames(time=750) # Don't let autogain run every long.
+sensor.set_auto_gain(False) # Turn off autogain.
+sensor.set_auto_exposure(False) # Turn off whitebalance.
+
+# Load cifar10 network (You can get the network from OpenMV IDE).
+net = nn.load('/cifar10.network')
+# Faster, smaller and less accurate.
+# net = nn.load('/cifar10_fast.network')
+labels = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
+
+clock = time.clock()
+while(True):
+ clock.tick()
+
+ img = sensor.snapshot()
+
+ # net.search() will search an roi in the image for the network (or the whole image if the roi is not
+ # specified). At each location to look in the image if one of the classifier outputs is larger than
+ # threshold the location and label will be stored in an object list and returned. At each scale the
+ # detection window is moved around in the ROI using x_overlap (0-1) and y_overlap (0-1) as a guide.
+ # If you set the overlap to 0.5 then each detection window will overlap the previous one by 50%. Note
+ # the computational work load goes WAY up the more overlap. Finally, for mult-scale matching after
+ # sliding the network around in the x/y dimensions the detection window will shrink by scale_mul (0-1)
+ # down to min_scale (0-1). For example, if scale_mul is 0.5 the detection window will shrink by 50%.
+ # Note that at a lower scale there's even more area to search if x_overlap and y_overlap are small...
+ # contrast_threshold skips running the CNN in areas that are flat.
+
+ # Setting x_overlap=-1 forces the window to stay centered in the ROI in the x direction always. If
+ # y_overlap is not -1 the method will search in all vertical positions.
+
+ # Setting y_overlap=-1 forces the window to stay centered in the ROI in the y direction always. If
+ # x_overlap is not -1 the method will serach in all horizontal positions.
+
+ for obj in net.search(img, threshold=0.6, min_scale=0.4, scale_mul=0.8, \
+ x_overlap=-1, y_overlap=-1, contrast_threshold=0.5):
+ print("Detected %s - Confidence %f%%" % (labels[obj.index()], obj.value()))
+ img.draw_rectangle(obj.rect(), color=(255, 0, 0))
+ print(clock.fps())
diff --git a/scripts/examples/25-Machine-Learning/nn_cifar10_search_whole_window.py b/scripts/examples/25-Machine-Learning/nn_cifar10_search_whole_window.py
new file mode 100644
index 000000000..c95acff19
--- /dev/null
+++ b/scripts/examples/25-Machine-Learning/nn_cifar10_search_whole_window.py
@@ -0,0 +1,47 @@
+# CIFAR-10 Search Whole Window Example
+#
+# CIFAR is a convolutional nueral network designed to classify it's field of view into several
+# different object types and works on RGB video data.
+#
+# In this example we slide the LeNet detector window over the image and get a list of activations
+# where there might be an object. Note that use a CNN with a sliding window is extremely compute
+# expensive so for an exhaustive search do not expect the CNN to be real-time.
+
+import sensor, image, time, os, nn
+
+sensor.reset() # Reset and initialize the sensor.
+sensor.set_pixformat(sensor.RGB565) # Set pixel format to RGB565 (or GRAYSCALE)
+sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
+sensor.set_windowing((128, 128)) # Set 128x128 window.
+sensor.skip_frames(time=750) # Don't let autogain run every long.
+sensor.set_auto_gain(False) # Turn off autogain.
+sensor.set_auto_exposure(False) # Turn off whitebalance.
+
+# Load cifar10 network (You can get the network from OpenMV IDE).
+net = nn.load('/cifar10.network')
+# Faster, smaller and less accurate.
+# net = nn.load('/cifar10_fast.network')
+labels = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck']
+
+clock = time.clock()
+while(True):
+ clock.tick()
+
+ img = sensor.snapshot()
+
+ # net.search() will search an roi in the image for the network (or the whole image if the roi is not
+ # specified). At each location to look in the image if one of the classifier outputs is larger than
+ # threshold the location and label will be stored in an object list and returned. At each scale the
+ # detection window is moved around in the ROI using x_overlap (0-1) and y_overlap (0-1) as a guide.
+ # If you set the overlap to 0.5 then each detection window will overlap the previous one by 50%. Note
+ # the computational work load goes WAY up the more overlap. Finally, for mult-scale matching after
+ # sliding the network around in the x/y dimensions the detection window will shrink by scale_mul (0-1)
+ # down to min_scale (0-1). For example, if scale_mul is 0.5 the detection window will shrink by 50%.
+ # Note that at a lower scale there's even more area to search if x_overlap and y_overlap are small...
+ # contrast_threshold skips running the CNN in areas that are flat.
+
+ for obj in net.search(img, threshold=0.6, min_scale=0.5, scale_mul=0.5, \
+ x_overlap=0.5, y_overlap=0.5, contrast_threshold=0.5):
+ print("Detected %s - Confidence %f%%" % (labels[obj.index()], obj.value()))
+ img.draw_rectangle(obj.rect(), color=(255, 0, 0))
+ print(clock.fps())
diff --git a/scripts/examples/25-Machine-Learning/nn_lenet_search_just_center.py b/scripts/examples/25-Machine-Learning/nn_lenet_search_just_center.py
new file mode 100644
index 000000000..f844e65ad
--- /dev/null
+++ b/scripts/examples/25-Machine-Learning/nn_lenet_search_just_center.py
@@ -0,0 +1,51 @@
+# LeNet Search Just Center Example
+#
+# LeNet is a convolutional nueral network designed to classify it's field of view into digits 0-9.
+#
+# In this example we slide the LeNet detector window over the image and get a list of activations
+# where there might be an object. Note that use a CNN with a sliding window is extremely compute
+# expensive so for an exhaustive search do not expect the CNN to be real-time.
+
+import sensor, image, time, os, nn
+
+sensor.reset() # Reset and initialize the sensor.
+sensor.set_pixformat(sensor.GRAYSCALE) # Set pixel format to RGB565 (or GRAYSCALE)
+sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
+sensor.set_windowing((128, 128)) # Set 128x128 window.
+sensor.skip_frames(time=500) # Don't let autogain run every long.
+sensor.set_auto_gain(False) # Turn off autogain.
+sensor.set_auto_exposure(False) # Turn off whitebalance.
+
+# Load lenet network (You can get the network from OpenMV IDE).
+net = nn.load('/lenet.network')
+labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
+
+clock = time.clock()
+while(True):
+ clock.tick()
+
+ img = sensor.snapshot()
+ tmp_img = img.copy().binary([(150, 255)], invert=True)
+
+ # net.search() will search an roi in the image for the network (or the whole image if the roi is not
+ # specified). At each location to look in the image if one of the classifier outputs is larger than
+ # threshold the location and label will be stored in an object list and returned. At each scale the
+ # detection window is moved around in the ROI using x_overlap (0-1) and y_overlap (0-1) as a guide.
+ # If you set the overlap to 0.5 then each detection window will overlap the previous one by 50%. Note
+ # the computational work load goes WAY up the more overlap. Finally, for mult-scale matching after
+ # sliding the network around in the x/y dimensions the detection window will shrink by scale_mul (0-1)
+ # down to min_scale (0-1). For example, if scale_mul is 0.5 the detection window will shrink by 50%.
+ # Note that at a lower scale there's even more area to search if x_overlap and y_overlap are small...
+ # contrast_threshold skips running the CNN in areas that are flat.
+
+ # Setting x_overlap=-1 forces the window to stay centered in the ROI in the x direction always. If
+ # y_overlap is not -1 the method will search in all vertical positions.
+
+ # Setting y_overlap=-1 forces the window to stay centered in the ROI in the y direction always. If
+ # x_overlap is not -1 the method will serach in all horizontal positions.
+
+ for obj in net.search(tmp_img, threshold=0.8, min_scale=0.4, scale_mul=0.8, \
+ x_overlap=-1, y_overlap=-1, contrast_threshold=0.5):
+ print("Detected %s - Confidence %f%%" % (labels[obj.index()], obj.value()))
+ img.draw_rectangle(obj.rect())
+ print(clock.fps())
diff --git a/scripts/examples/25-Machine-Learning/nn_lenet_search_whole_window.py b/scripts/examples/25-Machine-Learning/nn_lenet_search_whole_window.py
new file mode 100644
index 000000000..0d215e599
--- /dev/null
+++ b/scripts/examples/25-Machine-Learning/nn_lenet_search_whole_window.py
@@ -0,0 +1,45 @@
+# LeNet Search Whole Window Example
+#
+# LeNet is a convolutional nueral network designed to classify it's field of view into digits 0-9.
+#
+# In this example we slide the LeNet detector window over the image and get a list of activations
+# where there might be an object. Note that use a CNN with a sliding window is extremely compute
+# expensive so for an exhaustive search do not expect the CNN to be real-time.
+
+import sensor, image, time, os, nn
+
+sensor.reset() # Reset and initialize the sensor.
+sensor.set_pixformat(sensor.GRAYSCALE) # Set pixel format to RGB565 (or GRAYSCALE)
+sensor.set_framesize(sensor.QVGA) # Set frame size to QVGA (320x240)
+sensor.set_windowing((128, 128)) # Set 128x128 window.
+sensor.skip_frames(time=500) # Don't let autogain run every long.
+sensor.set_auto_gain(False) # Turn off autogain.
+sensor.set_auto_exposure(False) # Turn off whitebalance.
+
+# Load lenet network (You can get the network from OpenMV IDE).
+net = nn.load('/lenet.network')
+labels = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
+
+clock = time.clock()
+while(True):
+ clock.tick()
+
+ img = sensor.snapshot()
+ tmp_img = img.copy().binary([(150, 255)], invert=True)
+
+ # net.search() will search an roi in the image for the network (or the whole image if the roi is not
+ # specified). At each location to look in the image if one of the classifier outputs is larger than
+ # threshold the location and label will be stored in an object list and returned. At each scale the
+ # detection window is moved around in the ROI using x_overlap (0-1) and y_overlap (0-1) as a guide.
+ # If you set the overlap to 0.5 then each detection window will overlap the previous one by 50%. Note
+ # the computational work load goes WAY up the more overlap. Finally, for mult-scale matching after
+ # sliding the network around in the x/y dimensions the detection window will shrink by scale_mul (0-1)
+ # down to min_scale (0-1). For example, if scale_mul is 0.5 the detection window will shrink by 50%.
+ # Note that at a lower scale there's even more area to search if x_overlap and y_overlap are small...
+ # contrast_threshold skips running the CNN in areas that are flat.
+
+ for obj in net.search(tmp_img, threshold=0.9, min_scale=0.5, scale_mul=0.5, \
+ x_overlap=0.5, y_overlap=0.5, contrast_threshold=0.5):
+ print("Detected %s - Confidence %f%%" % (labels[obj.index()], obj.value()))
+ img.draw_rectangle(obj.rect())
+ print(clock.fps())
diff --git a/src/cmsis/include/arm_nnfunctions.h b/src/cmsis/include/arm_nnfunctions.h
index 375b9b856..c4015eca4 100644
--- a/src/cmsis/include/arm_nnfunctions.h
+++ b/src/cmsis/include/arm_nnfunctions.h
@@ -427,6 +427,69 @@ extern "C"
q15_t * bufferA,
q7_t * bufferB);
+ /**
+ * @brief Fast Q15 convolution function (non-sqaure shape)
+ * @param[in] Im_in pointer to input tensor
+ * @param[in] dim_im_in_x input tensor dimention x
+ * @param[in] dim_im_in_y input tensor dimention y
+ * @param[in] ch_im_in number of input tensor channels
+ * @param[in] wt pointer to kernel weights
+ * @param[in] ch_im_out number of filters, i.e., output tensor channels
+ * @param[in] dim_kernel_x filter kernel size x
+ * @param[in] dim_kernel_y filter kernel size y
+ * @param[in] padding_x padding size x
+ * @param[in] padding_y padding size y
+ * @param[in] stride_x convolution stride x
+ * @param[in] stride_y convolution stride y
+ * @param[in] bias pointer to bias
+ * @param[in] bias_shift amount of left-shift for bias
+ * @param[in] out_shift amount of right-shift for output
+ * @param[in,out] Im_out pointer to output tensor
+ * @param[in] dim_im_out_x output tensor dimension x
+ * @param[in] dim_im_out_y output tensor dimension y
+ * @param[in,out] bufferA pointer to buffer space for input
+ * @param[in,out] bufferB pointer to buffer space for output
+ * @return The function returns either
+ * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.
+ *
+ * @details
+ *
+ * Buffer size:
+ *
+ * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+ *
+ * bufferB size: 0
+ *
+ * Input dimension constraints:
+ *
+ * ch_im_in is multiple of 2
+ *
+ * ch_im_out is multipe of 2
+ *
+ */
+
+ arm_status
+ arm_convolve_HWC_q15_fast_nonsquare(const q15_t * Im_in,
+ const uint16_t dim_im_in_x,
+ const uint16_t dim_im_in_y,
+ const uint16_t ch_im_in,
+ const q15_t * wt,
+ const uint16_t ch_im_out,
+ const uint16_t dim_kernel_x,
+ const uint16_t dim_kernel_y,
+ const uint16_t padding_x,
+ const uint16_t padding_y,
+ const uint16_t stride_x,
+ const uint16_t stride_y,
+ const q15_t * bias,
+ const uint16_t bias_shift,
+ const uint16_t out_shift,
+ q15_t * Im_out,
+ const uint16_t dim_im_out_x,
+ const uint16_t dim_im_out_y,
+ q15_t * bufferA,
+ q7_t * bufferB);
+
/**
* @brief Q7 depthwise separable convolution function
* @param[in] Im_in pointer to input tensor
diff --git a/src/cmsis/src/nn/ActivationFunctions/arm_nn_activations_q7.c b/src/cmsis/src/nn/ActivationFunctions/arm_nn_activations_q7.c
index 4236e38c8..1ca429fea 100644
--- a/src/cmsis/src/nn/ActivationFunctions/arm_nn_activations_q7.c
+++ b/src/cmsis/src/nn/ActivationFunctions/arm_nn_activations_q7.c
@@ -80,7 +80,7 @@ void arm_nn_activations_direct_q7(q7_t * data, uint16_t size, uint16_t int_width
while (i)
{
in = *pIn++;
- out = lookup_table[(uint8_t) in >> shift_size];
+ out = lookup_table[(uint8_t) (in >> shift_size)];
*pOut++ = out;
i--;
}
diff --git a/src/cmsis/src/nn/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c b/src/cmsis/src/nn/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c
new file mode 100644
index 000000000..027420224
--- /dev/null
+++ b/src/cmsis/src/nn/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c
@@ -0,0 +1,265 @@
+/*
+ * Copyright (C) 2010-2018 Arm Limited or its affiliates. All rights reserved.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the License); you may
+ * not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an AS IS BASIS, WITHOUT
+ * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/* ----------------------------------------------------------------------
+ * Project: CMSIS NN Library
+ * Title: arm_convolve_HWC_q15_fast.c
+ * Description: Fast Q15 version of convolution
+ *
+ * $Date: 24. May 2018
+ * $Revision: V.1.0.0
+ *
+ * Target Processor: Cortex-M cores
+ *
+ * -------------------------------------------------------------------- */
+
+#include "arm_math.h"
+#include "arm_nnfunctions.h"
+
+/**
+ * @ingroup groupNN
+ */
+
+/**
+ * @addtogroup NNConv
+ * @{
+ */
+
+ /**
+ * @brief Fast Q15 convolution function (non-sqaure shape)
+ * @param[in] Im_in pointer to input tensor
+ * @param[in] dim_im_in_x input tensor dimention x
+ * @param[in] dim_im_in_y input tensor dimention y
+ * @param[in] ch_im_in number of input tensor channels
+ * @param[in] wt pointer to kernel weights
+ * @param[in] ch_im_out number of filters, i.e., output tensor channels
+ * @param[in] dim_kernel_x filter kernel size x
+ * @param[in] dim_kernel_y filter kernel size y
+ * @param[in] padding_x padding size x
+ * @param[in] padding_y padding size y
+ * @param[in] stride_x convolution stride x
+ * @param[in] stride_y convolution stride y
+ * @param[in] bias pointer to bias
+ * @param[in] bias_shift amount of left-shift for bias
+ * @param[in] out_shift amount of right-shift for output
+ * @param[in,out] Im_out pointer to output tensor
+ * @param[in] dim_im_out_x output tensor dimension x
+ * @param[in] dim_im_out_y output tensor dimension y
+ * @param[in,out] bufferA pointer to buffer space for input
+ * @param[in,out] bufferB pointer to buffer space for output
+ * @return The function returns either
+ * ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.
+ *
+ * @details
+ *
+ * Buffer size:
+ *
+ * bufferA size: 2*ch_im_in*dim_kernel*dim_kernel
+ *
+ * bufferB size: 0
+ *
+ * Input dimension constraints:
+ *
+ * ch_im_in is multiple of 2
+ *
+ * ch_im_out is multipe of 2
+ *
+ */
+
+arm_status
+arm_convolve_HWC_q15_fast_nonsquare(const q15_t * Im_in,
+ const uint16_t dim_im_in_x,
+ const uint16_t dim_im_in_y,
+ const uint16_t ch_im_in,
+ const q15_t * wt,
+ const uint16_t ch_im_out,
+ const uint16_t dim_kernel_x,
+ const uint16_t dim_kernel_y,
+ const uint16_t padding_x,
+ const uint16_t padding_y,
+ const uint16_t stride_x,
+ const uint16_t stride_y,
+ const q15_t * bias,
+ const uint16_t bias_shift,
+ const uint16_t out_shift,
+ q15_t * Im_out,
+ const uint16_t dim_im_out_x,
+ const uint16_t dim_im_out_y,
+ q15_t * bufferA,
+ q7_t * bufferB)
+{
+
+#if defined (ARM_MATH_DSP)
+ int16_t i_out_y, i_out_x, i_ker_y, i_ker_x;
+
+ q15_t *pBuffer = bufferA;
+ q15_t *im_buffer = bufferA;
+ q15_t *pOut = Im_out;
+
+ if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
+ {
+ /* check if the input dimension meets the constraints */
+ return ARM_MATH_SIZE_MISMATCH;
+ }
+
+ /* Run the following code for Cortex-M4 and Cortex-M7 */
+
+ /* This part implements the im2col function */
+ for (i_out_y = 0; i_out_y < dim_im_out_y; i_out_y++)
+ {
+ for (i_out_x = 0; i_out_x < dim_im_out_x; i_out_x++)
+ {
+ for (i_ker_y = i_out_y * stride_y - padding_y; i_ker_y < i_out_y * stride_y - padding_y + dim_kernel_y; i_ker_y++)
+ {
+ for (i_ker_x = i_out_x * stride_x - padding_x; i_ker_x < i_out_x * stride_x - padding_x + dim_kernel_x; i_ker_x++)
+ {
+ if (i_ker_y < 0 || i_ker_y >= dim_im_in_y || i_ker_x < 0 || i_ker_x >= dim_im_in_x)
+ {
+ /* arm_fill_q15(0, pBuffer, ch_im_in); */
+ memset(pBuffer, 0, sizeof(q15_t)*ch_im_in);
+ } else
+ {
+ /* arm_copy_q15((q15_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, pBuffer, ch_im_in); */
+ memcpy(pBuffer, (q15_t *) Im_in + (i_ker_y * dim_im_in_x + i_ker_x) * ch_im_in, sizeof(q15_t)*ch_im_in);
+ }
+ pBuffer += ch_im_in;
+ }
+ }
+
+ if (i_out_x & 0x1)
+ {
+ int i;
+ /* initialize the matrix pointers for A */
+ const q15_t *pA = wt;
+
+ /* set up the second output pointers */
+ q15_t *pOut2 = pOut + ch_im_out;
+
+ /* this loop over rows in A */
+ for (i = 0; i < ch_im_out; i += 2)
+ {
+ /* setup pointers for B */
+ q15_t *pB = im_buffer;
+ const q15_t *pB2 = pB + ch_im_in * dim_kernel_y * dim_kernel_x;
+
+ /* aling the second pointer for A */
+ const q15_t *pA2 = pA + ch_im_in * dim_kernel_y * dim_kernel_x;
+
+ /* init the sum with bias */
+ q31_t sum = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+ q31_t sum2 = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+ q31_t sum3 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
+ q31_t sum4 = ((q31_t)bias[i + 1] << bias_shift) + NN_ROUND(out_shift);
+
+ uint16_t colCnt = ch_im_in * dim_kernel_y * dim_kernel_x >> 1;
+ /* accumulate over the vector */
+ while (colCnt)
+ {
+ q31_t inA1 = *__SIMD32(pA)++;
+ q31_t inB1 = *__SIMD32(pB)++;
+ q31_t inA2 = *__SIMD32(pA2)++;
+ q31_t inB2 = *__SIMD32(pB2)++;
+
+ sum = __SMLAD(inA1, inB1, sum);
+ sum2 = __SMLAD(inA1, inB2, sum2);
+ sum3 = __SMLAD(inA2, inB1, sum3);
+ sum4 = __SMLAD(inA2, inB2, sum4);
+
+ colCnt--;
+ } /* while over colCnt */
+ colCnt = ch_im_in * dim_kernel_y * dim_kernel_x & 0x1;
+ while (colCnt)
+ {
+ q15_t inA1 = *pA++;
+ q15_t inB1 = *pB++;
+ q15_t inA2 = *pA2++;
+ q15_t inB2 = *pB2++;
+
+ sum += inA1 * inB1;
+ sum2 += inA1 * inB2;
+ sum3 += inA2 * inB1;
+ sum4 += inA2 * inB2;
+ colCnt--;
+ } /* while over colCnt */
+ *pOut++ = (q15_t) __SSAT(sum >> out_shift, 16);
+ *pOut++ = (q15_t) __SSAT(sum3 >> out_shift, 16);
+ *pOut2++ = (q15_t) __SSAT(sum2 >> out_shift, 16);
+ *pOut2++ = (q15_t) __SSAT(sum4 >> out_shift, 16);
+
+ /* skip the row computed with A2 */
+ pA += ch_im_in * dim_kernel_y * dim_kernel_x;
+ } /* for over ch_im_out */
+
+ pOut += ch_im_out;
+ /* counter reset */
+ pBuffer = im_buffer;
+ }
+ }
+ }
+
+#else
+ /* Run the following code as reference implementation for Cortex-M0 and Cortex-M3 */
+ uint16_t i, j, k, l, m, n;
+ int conv_out;
+ signed char in_row, in_col;
+
+ if (ch_im_in % 2 != 0 || ch_im_out % 2 != 0)
+ {
+ /* check if the input dimension meets the constraints */
+ return ARM_MATH_SIZE_MISMATCH;
+ }
+
+ for (i = 0; i < ch_im_out; i++)
+ {
+ for (j = 0; j < dim_im_out_y; j++)
+ {
+ for (k = 0; k < dim_im_out_x; k++)
+ {
+ conv_out = ((q31_t)bias[i] << bias_shift) + NN_ROUND(out_shift);
+ for (m = 0; m < dim_kernel_y; m++)
+ {
+ for (n = 0; n < dim_kernel_x; n++)
+ {
+ in_row = stride_y * j + m - padding_y;
+ in_col = stride_x * k + n - padding_x;
+ if (in_row >= 0 && in_col >= 0 && in_row < dim_im_in_y && in_col < dim_im_in_x)
+ {
+ for (l = 0; l < ch_im_in; l++)
+ {
+ conv_out +=
+ Im_in[(in_row * dim_im_in_x + in_col) * ch_im_in +
+ l] * wt[i * ch_im_in * dim_kernel_x * dim_kernel_y + (m * dim_kernel_x +
+ n) * ch_im_in + l];
+ }
+ }
+ }
+ }
+ Im_out[i + (j * dim_im_out_x + k) * ch_im_out] = (q15_t) __SSAT((conv_out >> out_shift), 16);
+ }
+ }
+ }
+
+#endif /* ARM_MATH_DSP */
+
+ /* Return to application */
+ return ARM_MATH_SUCCESS;
+}
+
+/**
+ * @} end of NNConv group
+ */
diff --git a/src/omv/nn/nn.c b/src/omv/nn/nn.c
index a1892d380..e89dae207 100644
--- a/src/omv/nn/nn.c
+++ b/src/omv/nn/nn.c
@@ -273,19 +273,27 @@ error:
return res;
}
-void nn_transform_input(data_layer_t *data_layer, image_t *img, q7_t *input_data)
+#ifndef __SSAT
+#define __SSAT(a, b) ({ __typeof__ (a) _a = (a); \
+ __typeof__ (b) _b = (b); \
+ _b = 1 << (_b - 1); \
+ _a = _a < (_b - 1) ? _a : (_b - 1); \
+ _a > (-_b) ? _a : (-_b); })
+#endif
+
+void nn_transform_input(data_layer_t *data_layer, image_t *img, q7_t *input_data, rectangle_t *roi)
{
int input_scale = data_layer->scale;
- // scale, convert and normalize input image.
- int x_ratio = (int)((img->w<<16)/data_layer->w)+1;
- int y_ratio = (int)((img->h<<16)/data_layer->h)+1;
+ // Scale, convert and normalize input image.
+ int x_ratio = (int)((roi->w<<16)/data_layer->w)+1;
+ int y_ratio = (int)((roi->h<<16)/data_layer->h)+1;
- if (img->bpp == 2 && data_layer->c == 3) { //RGB565 to RGB888
+ if ((img->bpp == 2) && (data_layer->c == 3)) { // RGB565 to RGB888
for (int y=0, i=0; yh; y++) {
int sy = (y*y_ratio)>>16;
for (int x=0; xw; x++, i+=3) {
int sx = (x*x_ratio)>>16;
- uint16_t p = IM_GET_RGB565_PIXEL(img, sx, sy);
+ uint16_t p = IM_GET_RGB565_PIXEL(img, sx+roi->x, sy+roi->y);
input_data[i+0] = (q7_t)__SSAT((((((int) COLOR_RGB565_TO_R8(p))
- (int) data_layer->r_mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
input_data[i+1] = (q7_t)__SSAT((((((int) COLOR_RGB565_TO_G8(p))
@@ -294,17 +302,17 @@ void nn_transform_input(data_layer_t *data_layer, image_t *img, q7_t *input_data
- (int) data_layer->b_mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
}
}
- } else if (img->bpp == 2 && data_layer->c == 1) { //RGB565 to GS
+ } else if ((img->bpp == 2) && (data_layer->c == 1)) { // RGB565 to GS
for (int y=0, i=0; yh; y++) {
int sy = (y*y_ratio)>>16;
for (int x=0; xw; x++, i++) {
int sx = (x*x_ratio)>>16;
- uint16_t p = IM_GET_RGB565_PIXEL(img, sx, sy);
+ uint16_t p = IM_GET_RGB565_PIXEL(img, sx+roi->x, sy+roi->y);
input_data[i] = (q7_t)__SSAT((((((int) COLOR_RGB565_TO_GRAYSCALE(p))
- (int) data_layer->r_mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
}
}
- } else if (img->bpp == 1 && data_layer->c == 3) { //GS to RGB88
+ } else if ((img->bpp == 1) && (data_layer->c == 3)) { // GS to RGB88
int mean = (int) ((0.30f * data_layer->r_mean) +
(0.59f * data_layer->g_mean) +
(0.11f * data_layer->b_mean));
@@ -312,25 +320,48 @@ void nn_transform_input(data_layer_t *data_layer, image_t *img, q7_t *input_data
int sy = (y*y_ratio)>>16;
for (int x=0; xw; x++, i+=3) {
int sx = (x*x_ratio)>>16;
- int p = (int) IMAGE_GET_GRAYSCALE_PIXEL(img, sx, sy);
+ int p = (int) IMAGE_GET_GRAYSCALE_PIXEL(img, sx+roi->x, sy+roi->y);
input_data[i+0] = (q7_t)__SSAT((((p - (int) mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
input_data[i+1] = (q7_t)__SSAT((((p - (int) mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
input_data[i+2] = (q7_t)__SSAT((((p - (int) mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
}
}
- } else if (img->bpp == 1 && data_layer->c == 1) { //GS to GS
+ } else if ((img->bpp == 1) && (data_layer->c == 1)) { // GS to GS
for (int y=0, i=0; yh; y++) {
int sy = (y*y_ratio)>>16;
for (int x=0; xw; x++, i++) {
int sx = (x*x_ratio)>>16;
- int p = (int) IMAGE_GET_GRAYSCALE_PIXEL(img, sx, sy);
+ int p = (int) IMAGE_GET_GRAYSCALE_PIXEL(img, sx+roi->x, sy+roi->y);
+ input_data[i] = (q7_t)__SSAT((((p - (int) data_layer->r_mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
+ }
+ }
+ } else if ((img->bpp == 0) && (data_layer->c == 3)) { // BINARY to RGB88
+ int mean = (int) ((0.30f * data_layer->r_mean) +
+ (0.59f * data_layer->g_mean) +
+ (0.11f * data_layer->b_mean));
+ for (int y=0, i=0; yh; y++) {
+ int sy = (y*y_ratio)>>16;
+ for (int x=0; xw; x++, i+=3) {
+ int sx = (x*x_ratio)>>16;
+ int p = (int) COLOR_BINARY_TO_GRAYSCALE(IMAGE_GET_BINARY_PIXEL(img, sx+roi->x, sy+roi->y));
+ input_data[i+0] = (q7_t)__SSAT((((p - (int) mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
+ input_data[i+1] = (q7_t)__SSAT((((p - (int) mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
+ input_data[i+2] = (q7_t)__SSAT((((p - (int) mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
+ }
+ }
+ } else if ((img->bpp == 0) && (data_layer->c == 1)) { // BINARY to GS
+ for (int y=0, i=0; yh; y++) {
+ int sy = (y*y_ratio)>>16;
+ for (int x=0; xw; x++, i++) {
+ int sx = (x*x_ratio)>>16;
+ int p = (int) COLOR_BINARY_TO_GRAYSCALE(IMAGE_GET_BINARY_PIXEL(img, sx+roi->x, sy+roi->y));
input_data[i] = (q7_t)__SSAT((((p - (int) data_layer->r_mean)<<7) + (1<<(input_scale-1))) >> input_scale, 8);
}
}
}
}
-int nn_run_network(nn_t *net, image_t *img, bool softmax)
+int nn_run_network(nn_t *net, image_t *img, rectangle_t *roi, bool softmax)
{
uint32_t layer_idx = 0;
layer_t *layer = net->layers;
@@ -360,7 +391,7 @@ int nn_run_network(nn_t *net, image_t *img, bool softmax)
case LAYER_TYPE_DATA: {
data_layer_t *data_layer = (data_layer_t *) layer;
input_data = fb_alloc(data_layer->c * data_layer->h * data_layer->w);
- nn_transform_input(data_layer, img, input_data);
+ nn_transform_input(data_layer, img, input_data, roi);
// Set image data as input buffer for the next layer.
input_buffer = input_data;
output_buffer = buffer1;
diff --git a/src/omv/nn/nn.h b/src/omv/nn/nn.h
index 116ae82ce..0ee115240 100644
--- a/src/omv/nn/nn.h
+++ b/src/omv/nn/nn.h
@@ -100,6 +100,6 @@ typedef void (*pool_func_t)(q7_t * Im_in, const uint16_t dim_im_in, const uint16
int nn_dump_network(nn_t *net);
int nn_load_network(nn_t *net, const char *path);
-int nn_run_network(nn_t *net, image_t *img, bool softmax);
+int nn_run_network(nn_t *net, image_t *img, rectangle_t *roi, bool softmax);
int nn_dry_run_network(nn_t *net, image_t *img, bool softmax);
#endif //#define __CNN_H__
diff --git a/src/omv/py/py_nn.c b/src/omv/py/py_nn.c
index b07c7778a..d08c2dd68 100644
--- a/src/omv/py/py_nn.c
+++ b/src/omv/py/py_nn.c
@@ -1,22 +1,15 @@
-/*
- * This file is part of the OpenMV project.
- * Copyright (c) 2013/2014 Ibrahim Abdelkader
+/* This file is part of the OpenMV project.
+ * Copyright (c) 2013-2018 Ibrahim Abdelkader & Kwabena W. Agyeman
* This work is licensed under the MIT license, see the file LICENSE for details.
- *
- * NN module.
- *
*/
-#include "mp.h"
+#include
#include "nn.h"
-#include "imlib.h"
-#include "xalloc.h"
-#include "py_image.h"
#include "py_helper.h"
-#include "py_assert.h"
+#include "py_image.h"
#include "omv_boardconfig.h"
-#include
#ifdef IMLIB_ENABLE_CNN
+
static const mp_obj_type_t py_net_type;
typedef struct _py_net_obj_t {
@@ -24,31 +17,10 @@ typedef struct _py_net_obj_t {
nn_t _cobj;
} py_net_obj_t;
-void *py_net_cobj(mp_obj_t net)
+void *py_net_cobj(mp_obj_t net_obj)
{
- PY_ASSERT_TYPE(net, &py_net_type);
- return &((py_net_obj_t *)net)->_cobj;
-}
-
-STATIC mp_obj_t py_net_forward(uint n_args, const mp_obj_t *args, mp_map_t *kw_args)
-{
- nn_t *net = py_net_cobj(args[0]);
- image_t *img = py_helper_arg_to_image_mutable(args[1]);
-
- mp_obj_t output_list = mp_obj_new_list(0, NULL);
- bool softmax = py_helper_keyword_int(n_args, args, 2, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_softmax), false);
- bool dry_run = py_helper_keyword_int(n_args, args, 3, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_dry_run), false);
-
- if (dry_run == false) {
- nn_run_network(net, img, softmax);
- } else {
- nn_dry_run_network(net, img, softmax);
- }
-
- for (int i=0; ioutput_size; i++) {
- mp_obj_list_append(output_list, mp_obj_new_int(net->output_data[i]));
- }
- return output_list;
+ PY_ASSERT_TYPE(net_obj, &py_net_type);
+ return &((py_net_obj_t *)net_obj)->_cobj;
}
STATIC void py_net_print(const mp_print_t *print, mp_obj_t self_in, mp_print_kind_t kind)
@@ -57,12 +29,343 @@ STATIC void py_net_print(const mp_print_t *print, mp_obj_t self_in, mp_print_kin
nn_dump_network(py_net_cobj(self));
}
+STATIC mp_obj_t py_net_forward(uint n_args, const mp_obj_t *args, mp_map_t *kw_args)
+{
+ nn_t *net = py_net_cobj(args[0]);
+ image_t *img = py_helper_arg_to_image_mutable(args[1]);
+
+ rectangle_t roi;
+ py_helper_keyword_rectangle_roi(img, n_args, args, 2, kw_args, &roi);
+
+ bool softmax = py_helper_keyword_int(n_args, args, 3, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_softmax), false);
+ bool dry_run = py_helper_keyword_int(n_args, args, 4, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_dry_run), false);
+
+ mp_obj_t output_list = mp_obj_new_list(0, NULL);
+ fb_alloc_mark();
+
+ if (dry_run == false) {
+ nn_run_network(net, img, &roi, softmax);
+ } else {
+ nn_dry_run_network(net, img, softmax);
+ }
+
+ for (int i=0; ioutput_size; i++) {
+ mp_obj_list_append(output_list, mp_obj_new_int(net->output_data[i]));
+ }
+
+ fb_alloc_free_till_mark();
+ return output_list;
+}
STATIC MP_DEFINE_CONST_FUN_OBJ_KW(py_net_forward_obj, 2, py_net_forward);
+// NN Class Object
+#define py_nn_class_obj_size 6
+typedef struct py_nn_class_obj {
+ mp_obj_base_t base;
+ mp_obj_t x, y, w, h, index, value;
+} py_nn_class_obj_t;
-static const mp_map_elem_t locals_dict_table[] = {
- { MP_OBJ_NEW_QSTR(MP_QSTR_forward), (mp_obj_t)&py_net_forward_obj},
- { NULL, NULL },
+static void py_nn_class_print(const mp_print_t *print, mp_obj_t self_in, mp_print_kind_t kind)
+{
+ py_nn_class_obj_t *self = self_in;
+ mp_printf(print,
+ "{\"x\":%d, \"y\":%d, \"w\":%d, \"h\":%d, \"index\":%d, \"value\":%f}",
+ mp_obj_get_int(self->x),
+ mp_obj_get_int(self->y),
+ mp_obj_get_int(self->w),
+ mp_obj_get_int(self->h),
+ mp_obj_get_int(self->index),
+ (double) mp_obj_get_float(self->value));
+}
+
+static mp_obj_t py_nn_class_subscr(mp_obj_t self_in, mp_obj_t index, mp_obj_t value)
+{
+ if (value == MP_OBJ_SENTINEL) { // load
+ py_nn_class_obj_t *self = self_in;
+ if (MP_OBJ_IS_TYPE(index, &mp_type_slice)) {
+ mp_bound_slice_t slice;
+ if (!mp_seq_get_fast_slice_indexes(py_nn_class_obj_size, index, &slice)) {
+ nlr_raise(mp_obj_new_exception_msg(&mp_type_OSError, "only slices with step=1 (aka None) are supported"));
+ }
+ mp_obj_tuple_t *result = mp_obj_new_tuple(slice.stop - slice.start, NULL);
+ mp_seq_copy(result->items, &(self->x) + slice.start, result->len, mp_obj_t);
+ return result;
+ }
+ switch (mp_get_index(self->base.type, py_nn_class_obj_size, index, false)) {
+ case 0: return self->x;
+ case 1: return self->y;
+ case 2: return self->w;
+ case 3: return self->h;
+ case 4: return self->index;
+ case 5: return self->value;
+ }
+ }
+ return MP_OBJ_NULL; // op not supported
+}
+
+mp_obj_t py_nn_class_rect(mp_obj_t self_in)
+{
+ return mp_obj_new_tuple(4, (mp_obj_t []) {((py_nn_class_obj_t *) self_in)->x,
+ ((py_nn_class_obj_t *) self_in)->y,
+ ((py_nn_class_obj_t *) self_in)->w,
+ ((py_nn_class_obj_t *) self_in)->h});
+}
+
+mp_obj_t py_nn_class_x(mp_obj_t self_in) { return ((py_nn_class_obj_t *) self_in)->x; }
+mp_obj_t py_nn_class_y(mp_obj_t self_in) { return ((py_nn_class_obj_t *) self_in)->y; }
+mp_obj_t py_nn_class_w(mp_obj_t self_in) { return ((py_nn_class_obj_t *) self_in)->w; }
+mp_obj_t py_nn_class_h(mp_obj_t self_in) { return ((py_nn_class_obj_t *) self_in)->h; }
+mp_obj_t py_nn_class_index(mp_obj_t self_in) { return ((py_nn_class_obj_t *) self_in)->index; }
+mp_obj_t py_nn_class_value(mp_obj_t self_in) { return ((py_nn_class_obj_t *) self_in)->value; }
+
+STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_class_rect_obj, py_nn_class_rect);
+STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_class_x_obj, py_nn_class_x);
+STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_class_y_obj, py_nn_class_y);
+STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_class_w_obj, py_nn_class_w);
+STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_class_h_obj, py_nn_class_h);
+STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_class_index_obj, py_nn_class_index);
+STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_class_value_obj, py_nn_class_value);
+
+STATIC const mp_rom_map_elem_t py_nn_class_locals_dict_table[] = {
+ { MP_ROM_QSTR(MP_QSTR_rect), MP_ROM_PTR(&py_nn_class_rect_obj) },
+ { MP_ROM_QSTR(MP_QSTR_x), MP_ROM_PTR(&py_nn_class_x_obj) },
+ { MP_ROM_QSTR(MP_QSTR_y), MP_ROM_PTR(&py_nn_class_y_obj) },
+ { MP_ROM_QSTR(MP_QSTR_w), MP_ROM_PTR(&py_nn_class_w_obj) },
+ { MP_ROM_QSTR(MP_QSTR_h), MP_ROM_PTR(&py_nn_class_h_obj) },
+ { MP_ROM_QSTR(MP_QSTR_index), MP_ROM_PTR(&py_nn_class_index_obj) },
+ { MP_ROM_QSTR(MP_QSTR_value), MP_ROM_PTR(&py_nn_class_value_obj) }
+};
+
+STATIC MP_DEFINE_CONST_DICT(py_nn_class_locals_dict, py_nn_class_locals_dict_table);
+
+static const mp_obj_type_t py_nn_class_type = {
+ { &mp_type_type },
+ .name = MP_QSTR_nn_class,
+ .print = py_nn_class_print,
+ .subscr = py_nn_class_subscr,
+ .locals_dict = (mp_obj_t) &py_nn_class_locals_dict
+};
+
+typedef struct py_nn_class_obj_list_lnk_data {
+ rectangle_t rect;
+ int index;
+ float value;
+ int merge_number;
+} py_nn_class_obj_list_lnk_data_t;
+
+STATIC mp_obj_t py_net_search(uint n_args, const mp_obj_t *args, mp_map_t *kw_args)
+{
+ nn_t *arg_net = py_net_cobj(args[0]);
+ image_t *arg_img = py_helper_arg_to_image_mutable(args[1]);
+
+ rectangle_t roi;
+ py_helper_keyword_rectangle_roi(arg_img, n_args, args, 2, kw_args, &roi);
+
+ float arg_threshold = py_helper_keyword_float(n_args, args, 3, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_threshold), 0.6);
+ PY_ASSERT_TRUE_MSG((0 <= arg_threshold) && (arg_threshold <= 1), "0 <= threshold <= 1");
+
+ float arg_min_scale = py_helper_keyword_float(n_args, args, 4, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_min_scale), 1.0);
+ PY_ASSERT_TRUE_MSG((0 < arg_min_scale) && (arg_min_scale <= 1), "0 < min_scale <= 1");
+
+ float arg_scale_mul = py_helper_keyword_float(n_args, args, 5, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_scale_mul), 0.5);
+ PY_ASSERT_TRUE_MSG((0 <= arg_scale_mul) && (arg_scale_mul < 1), "0 <= scale_mul < 1");
+
+ float arg_x_overlap = py_helper_keyword_float(n_args, args, 6, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_x_overlap), 0);
+ PY_ASSERT_TRUE_MSG(((0 <= arg_x_overlap) && (arg_x_overlap < 1)) || (arg_x_overlap == -1), "0 <= x_overlap < 1");
+
+ float arg_y_overlap = py_helper_keyword_float(n_args, args, 7, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_y_overlap), 0);
+ PY_ASSERT_TRUE_MSG(((0 <= arg_y_overlap) && (arg_y_overlap < 1)) || (arg_y_overlap == -1), "0 <= y_overlap < 1");
+
+ float arg_contrast_threshold = py_helper_keyword_float(n_args, args, 8, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_contrast_threshold), 1);
+ PY_ASSERT_TRUE_MSG(0 <= arg_contrast_threshold, "0 <= contrast_threshold");
+
+ bool softmax = py_helper_keyword_int(n_args, args, 9, kw_args, MP_OBJ_NEW_QSTR(MP_QSTR_softmax), false);
+
+ list_t out;
+ list_init(&out, sizeof(py_nn_class_obj_list_lnk_data_t));
+
+ fb_alloc_mark();
+
+ for (float scale = 1; scale >= arg_min_scale; scale *= arg_scale_mul) {
+ // Either provide a subtle offset to center multiple detection windows or center the only detection window.
+ for (int y = roi.y + ((arg_y_overlap != -1) ? (fmodf(roi.h, (roi.h * scale)) / 2) : ((roi.h - (roi.h * scale)) / 2));
+ // Finish when the detection window is outside of the ROI.
+ (y + (roi.h * scale)) <= (roi.y + roi.h);
+ // Step by an overlap amount accounting for scale or just terminate after one iteration.
+ y += ((arg_y_overlap != -1) ? (roi.h * scale * (1 - arg_y_overlap)) : roi.h)) {
+ // Either provide a subtle offset to center multiple detection windows or center the only detection window.
+ for (int x = roi.x + ((arg_x_overlap != -1) ? (fmodf(roi.w, (roi.w * scale)) / 2) : ((roi.w - (roi.w * scale)) / 2));
+ // Finish when the detection window is outside of the ROI.
+ (x + (roi.w * scale)) <= (roi.x + roi.w);
+ // Step by an overlap amount accounting for scale or just terminate after one iteration.
+ x += ((arg_x_overlap != -1) ? (roi.w * scale * (1 - arg_x_overlap)) : roi.w)) {
+ rectangle_t new_roi;
+ rectangle_init(&new_roi, x, y, roi.w * scale, roi.h * scale);
+ if (rectangle_overlap(&roi, &new_roi)) {
+
+ int sum = 0;
+ int sum_2 = 0;
+ for (int b = new_roi.y, bb = new_roi.y + new_roi.h, bbb = fast_sqrtf(new_roi.h); b < bb; b += bbb) {
+ for (int a = new_roi.x, aa = new_roi.x + new_roi.w, aaa = fast_sqrtf(new_roi.w); a < aa; a += aaa) {
+ switch(arg_img->bpp) {
+ case IMAGE_BPP_BINARY: {
+ int pixel = COLOR_BINARY_TO_GRAYSCALE(IMAGE_GET_BINARY_PIXEL(arg_img, a, b));
+ sum += pixel;
+ sum_2 += pixel * pixel;
+ }
+ case IMAGE_BPP_GRAYSCALE: {
+ int pixel = IMAGE_GET_GRAYSCALE_PIXEL(arg_img, a, b);
+ sum += pixel;
+ sum_2 += pixel * pixel;
+ }
+ case IMAGE_BPP_RGB565: {
+ int pixel = COLOR_RGB565_TO_GRAYSCALE(IMAGE_GET_RGB565_PIXEL(arg_img, a, b));
+ sum += pixel;
+ sum_2 += pixel * pixel;
+ }
+ }
+ }
+ }
+
+ int area = new_roi.w * new_roi.h;
+ int mean = sum / area;
+ int variance = (sum_2 / area) - (mean * mean);
+
+ if (fast_sqrtf(variance) >= arg_contrast_threshold) { // Skip flat regions...
+ nn_run_network(arg_net, arg_img, &new_roi, softmax);
+
+ int max_index = -1;
+ float max_value = -1;
+ for (int i=0; ioutput_size; i++) {
+ float value = ((float) (arg_net->output_data[i] + 128)) / 255;
+ if ((value >= arg_threshold) && (value > max_value)) {
+ max_index = i;
+ max_value = value;
+ }
+ }
+
+ if (max_index != -1) {
+ py_nn_class_obj_list_lnk_data_t lnk_data;
+ lnk_data.rect.x = new_roi.x;
+ lnk_data.rect.y = new_roi.y;
+ lnk_data.rect.w = new_roi.w;
+ lnk_data.rect.h = new_roi.h;
+ lnk_data.index = max_index;
+ lnk_data.value = max_value;
+ lnk_data.merge_number = 1;
+ list_push_back(&out, &lnk_data);
+ }
+ }
+ }
+ }
+ }
+ }
+
+ fb_alloc_free_till_mark();
+
+ // Merge all overlapping and same detections and average them.
+
+ for (;;) {
+ bool merge_occured = false;
+
+ list_t out_temp;
+ list_init(&out_temp, sizeof(py_nn_class_obj_list_lnk_data_t));
+
+ while (list_size(&out)) {
+ py_nn_class_obj_list_lnk_data_t lnk_data;
+ list_pop_front(&out, &lnk_data);
+
+ for (size_t k = 0, l = list_size(&out); k < l; k++) {
+ py_nn_class_obj_list_lnk_data_t tmp_data;
+ list_pop_front(&out, &tmp_data);
+
+ if ((lnk_data.index == tmp_data.index)
+ && rectangle_overlap(&(lnk_data.rect), &(tmp_data.rect))) {
+ lnk_data.rect.x = ((lnk_data.rect.x * lnk_data.merge_number) + tmp_data.rect.x) / (lnk_data.merge_number + 1);
+ lnk_data.rect.y = ((lnk_data.rect.y * lnk_data.merge_number) + tmp_data.rect.y) / (lnk_data.merge_number + 1);
+ lnk_data.rect.w = ((lnk_data.rect.w * lnk_data.merge_number) + tmp_data.rect.w) / (lnk_data.merge_number + 1);
+ lnk_data.rect.h = ((lnk_data.rect.h * lnk_data.merge_number) + tmp_data.rect.h) / (lnk_data.merge_number + 1);
+ lnk_data.value = ((lnk_data.value * lnk_data.merge_number) + tmp_data.value) / (lnk_data.merge_number + 1);
+ lnk_data.merge_number += 1;
+ merge_occured = true;
+ } else {
+ list_push_back(&out, &tmp_data);
+ }
+ }
+
+ list_push_back(&out_temp, &lnk_data);
+ }
+
+ list_copy(&out, &out_temp);
+
+ if (!merge_occured) {
+ break;
+ }
+ }
+
+ // Determine the winner between overlapping different class detections.
+
+ for (;;) {
+ bool merge_occured = false;
+
+ list_t out_temp;
+ list_init(&out_temp, sizeof(py_nn_class_obj_list_lnk_data_t));
+
+ while (list_size(&out)) {
+ py_nn_class_obj_list_lnk_data_t lnk_data;
+ list_pop_front(&out, &lnk_data);
+
+ for (size_t k = 0, l = list_size(&out); k < l; k++) {
+ py_nn_class_obj_list_lnk_data_t tmp_data;
+ list_pop_front(&out, &tmp_data);
+
+ if ((lnk_data.index != tmp_data.index)
+ && rectangle_overlap(&(lnk_data.rect), &(tmp_data.rect))) {
+ if (tmp_data.value > lnk_data.value) {
+ memcpy(&lnk_data, &tmp_data, sizeof(py_nn_class_obj_list_lnk_data_t));
+ }
+
+ merge_occured = true;
+ } else {
+ list_push_back(&out, &tmp_data);
+ }
+ }
+
+ list_push_back(&out_temp, &lnk_data);
+ }
+
+ list_copy(&out, &out_temp);
+
+ if (!merge_occured) {
+ break;
+ }
+ }
+
+ mp_obj_list_t *objects_list = mp_obj_new_list(list_size(&out), NULL);
+
+ for (size_t i = 0; list_size(&out); i++) {
+ py_nn_class_obj_list_lnk_data_t lnk_data;
+ list_pop_front(&out, &lnk_data);
+
+ py_nn_class_obj_t *o = m_new_obj(py_nn_class_obj_t);
+ o->base.type = &py_nn_class_type;
+ o->x = mp_obj_new_int(lnk_data.rect.x);
+ o->y = mp_obj_new_int(lnk_data.rect.y);
+ o->w = mp_obj_new_int(lnk_data.rect.w);
+ o->h = mp_obj_new_int(lnk_data.rect.h);
+ o->index = mp_obj_new_int(lnk_data.index);
+ o->value = mp_obj_new_float(lnk_data.value);
+
+ objects_list->items[i] = o;
+ }
+
+ return objects_list;
+}
+STATIC MP_DEFINE_CONST_FUN_OBJ_KW(py_net_search_obj, 2, py_net_search);
+
+STATIC const mp_rom_map_elem_t locals_dict_table[] = {
+ { MP_ROM_QSTR(MP_QSTR_forward), MP_ROM_PTR(&py_net_forward_obj) },
+ { MP_ROM_QSTR(MP_QSTR_search), MP_ROM_PTR(&py_net_search_obj) }
};
STATIC MP_DEFINE_CONST_DICT(locals_dict, locals_dict_table);
@@ -71,28 +374,28 @@ static const mp_obj_type_t py_net_type = {
{ &mp_type_type },
.name = MP_QSTR_Net,
.print = py_net_print,
- .locals_dict = (mp_obj_t)&locals_dict,
+ .locals_dict = (mp_obj_t) &locals_dict
};
static mp_obj_t py_nn_load(mp_obj_t path_obj)
{
- py_net_obj_t *net = NULL;
const char *path = mp_obj_str_get_str(path_obj);
- net = m_new_obj(py_net_obj_t);
+ py_net_obj_t *net = m_new_obj(py_net_obj_t);
net->base.type = &py_net_type;
nn_load_network(py_net_cobj(net), path);
return net;
}
STATIC MP_DEFINE_CONST_FUN_OBJ_1(py_nn_load_obj, py_nn_load);
+
#endif // IMLIB_ENABLE_CNN
-static const mp_map_elem_t globals_dict_table[] = {
- { MP_OBJ_NEW_QSTR(MP_QSTR___name__), MP_OBJ_NEW_QSTR(MP_QSTR_nn) },
+STATIC const mp_rom_map_elem_t globals_dict_table[] = {
+ { MP_ROM_QSTR(MP_QSTR___name__), MP_OBJ_NEW_QSTR(MP_QSTR_nn) },
#ifdef IMLIB_ENABLE_CNN
- { MP_OBJ_NEW_QSTR(MP_QSTR_load), (mp_obj_t)&py_nn_load_obj },
+ { MP_ROM_QSTR(MP_QSTR_load), MP_ROM_PTR(&py_nn_load_obj) },
#else
- { MP_OBJ_NEW_QSTR(MP_QSTR_load), (mp_obj_t)&py_func_unavailable_obj },
+ { MP_ROM_QSTR(MP_QSTR_load), MP_ROM_PTR(&py_func_unavailable_obj) }
#endif // IMLIB_ENABLE_CNN
};
@@ -100,5 +403,5 @@ STATIC MP_DEFINE_CONST_DICT(globals_dict, globals_dict_table);
const mp_obj_module_t nn_module = {
.base = { &mp_type_module },
- .globals = (mp_obj_t)&globals_dict,
+ .globals = (mp_obj_t) &globals_dict
};
diff --git a/src/omv/py/qstrdefsomv.h b/src/omv/py/qstrdefsomv.h
index 3e20fc615..1e23d74d3 100644
--- a/src/omv/py/qstrdefsomv.h
+++ b/src/omv/py/qstrdefsomv.h
@@ -138,10 +138,31 @@ Q(load)
// Net
Q(Net)
+
+// Forward
Q(forward)
Q(dry_run)
Q(softmax)
+// Search
+// duplicate Q(search)
+// duplicate Q(roi)
+// duplicate Q(threshold)
+Q(min_scale)
+Q(scale_mul)
+Q(x_overlap)
+Q(y_overlap)
+Q(contrast_threshold)
+// duplicate Q(softmax)
+// NN Class
+Q(nn_class)
+// duplicate Q(x)
+// duplicate Q(y)
+// duplicate Q(w)
+// duplicate Q(h)
+// duplicate Q(index)
+// duplicate Q(value)
+
// C/SIF Resolutions
Q(QQCIF)
Q(QCIF)