Compute Library
 21.02
NEHOGMultiDetection.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016-2020 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
25 
26 #include "arm_compute/core/Error.h"
34 
35 namespace arm_compute
36 {
38 
39 NEHOGMultiDetection::NEHOGMultiDetection(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
40  : _memory_group(std::move(memory_manager)),
41  _gradient_kernel(),
42  _orient_bin_kernel(),
43  _block_norm_kernel(),
44  _hog_detect_kernel(),
45  _non_maxima_kernel(),
46  _hog_space(),
47  _hog_norm_space(),
48  _detection_windows(),
49  _mag(),
50  _phase(),
51  _non_maxima_suppression(false),
52  _num_orient_bin_kernel(0),
53  _num_block_norm_kernel(0),
54  _num_hog_detect_kernel(0)
55 {
56 }
57 
58 void NEHOGMultiDetection::configure(ITensor *input, const IMultiHOG *multi_hog, IDetectionWindowArray *detection_windows, const ISize2DArray *detection_window_strides, BorderMode border_mode,
59  uint8_t constant_border_value, float threshold, bool non_maxima_suppression, float min_distance)
60 {
63  ARM_COMPUTE_ERROR_ON(nullptr == detection_windows);
64  ARM_COMPUTE_ERROR_ON(detection_window_strides->num_values() != multi_hog->num_models());
65 
66  const size_t width = input->info()->dimension(Window::DimX);
67  const size_t height = input->info()->dimension(Window::DimY);
68  const TensorShape &shape_img = input->info()->tensor_shape();
69  const size_t num_models = multi_hog->num_models();
70  PhaseType phase_type = multi_hog->model(0)->info()->phase_type();
71 
72  size_t prev_num_bins = multi_hog->model(0)->info()->num_bins();
73  Size2D prev_cell_size = multi_hog->model(0)->info()->cell_size();
74  Size2D prev_block_size = multi_hog->model(0)->info()->block_size();
75  Size2D prev_block_stride = multi_hog->model(0)->info()->block_stride();
76 
77  /* Check if NEHOGOrientationBinningKernel and NEHOGBlockNormalizationKernel kernels can be skipped for a specific HOG data-object
78  *
79  * 1) NEHOGOrientationBinningKernel and NEHOGBlockNormalizationKernel are skipped if the cell size and the number of bins don't change.
80  * Since "multi_hog" is sorted,it is enough to check the HOG descriptors at level "ith" and level "(i-1)th
81  * 2) NEHOGBlockNormalizationKernel is skipped if the cell size, the number of bins and block size do not change.
82  * Since "multi_hog" is sorted,it is enough to check the HOG descriptors at level "ith" and level "(i-1)th
83  *
84  * @note Since the orientation binning and block normalization kernels can be skipped, we need to keep track of the input to process for each kernel
85  * with "input_orient_bin", "input_hog_detect" and "input_block_norm"
86  */
87  std::vector<size_t> input_orient_bin;
88  std::vector<size_t> input_hog_detect;
89  std::vector<std::pair<size_t, size_t>> input_block_norm;
90 
91  input_orient_bin.push_back(0);
92  input_hog_detect.push_back(0);
93  input_block_norm.emplace_back(0, 0);
94 
95  for(size_t i = 1; i < num_models; ++i)
96  {
97  size_t cur_num_bins = multi_hog->model(i)->info()->num_bins();
98  Size2D cur_cell_size = multi_hog->model(i)->info()->cell_size();
99  Size2D cur_block_size = multi_hog->model(i)->info()->block_size();
100  Size2D cur_block_stride = multi_hog->model(i)->info()->block_stride();
101 
102  if((cur_num_bins != prev_num_bins) || (cur_cell_size.width != prev_cell_size.width) || (cur_cell_size.height != prev_cell_size.height))
103  {
104  prev_num_bins = cur_num_bins;
105  prev_cell_size = cur_cell_size;
106  prev_block_size = cur_block_size;
107  prev_block_stride = cur_block_stride;
108 
109  // Compute orientation binning and block normalization kernels. Update input to process
110  input_orient_bin.push_back(i);
111  input_block_norm.emplace_back(i, input_orient_bin.size() - 1);
112  }
113  else if((cur_block_size.width != prev_block_size.width) || (cur_block_size.height != prev_block_size.height) || (cur_block_stride.width != prev_block_stride.width)
114  || (cur_block_stride.height != prev_block_stride.height))
115  {
116  prev_block_size = cur_block_size;
117  prev_block_stride = cur_block_stride;
118 
119  // Compute block normalization kernel. Update input to process
120  input_block_norm.emplace_back(i, input_orient_bin.size() - 1);
121  }
122 
123  // Update input to process for hog detector kernel
124  input_hog_detect.push_back(input_block_norm.size() - 1);
125  }
126 
127  _detection_windows = detection_windows;
128  _non_maxima_suppression = non_maxima_suppression;
129  _num_orient_bin_kernel = input_orient_bin.size(); // Number of NEHOGOrientationBinningKernel kernels to compute
130  _num_block_norm_kernel = input_block_norm.size(); // Number of NEHOGBlockNormalizationKernel kernels to compute
131  _num_hog_detect_kernel = input_hog_detect.size(); // Number of NEHOGDetector functions to compute
132 
133  _orient_bin_kernel.clear();
134  _block_norm_kernel.clear();
135  _hog_detect_kernel.clear();
136  _hog_space.clear();
137  _hog_norm_space.clear();
138 
139  _orient_bin_kernel.resize(_num_orient_bin_kernel);
140  _block_norm_kernel.resize(_num_block_norm_kernel);
141  _hog_detect_kernel.resize(_num_hog_detect_kernel);
142  _hog_space.resize(_num_orient_bin_kernel);
143  _hog_norm_space.resize(_num_block_norm_kernel);
144  _non_maxima_kernel = CPPDetectionWindowNonMaximaSuppressionKernel();
145 
146  // Allocate tensors for magnitude and phase
147  TensorInfo info_mag(shape_img, Format::S16);
148  _mag.allocator()->init(info_mag);
149 
150  TensorInfo info_phase(shape_img, Format::U8);
151  _phase.allocator()->init(info_phase);
152 
153  // Manage intermediate buffers
154  _memory_group.manage(&_mag);
155  _memory_group.manage(&_phase);
156 
157  // Initialise gradient kernel
158  _gradient_kernel.configure(input, &_mag, &_phase, phase_type, border_mode, constant_border_value);
159 
160  // Configure NETensor for the HOG space and orientation binning kernel
161  for(size_t i = 0; i < _num_orient_bin_kernel; ++i)
162  {
163  const size_t idx_multi_hog = input_orient_bin[i];
164 
165  // Get the corresponding cell size and number of bins
166  const Size2D &cell = multi_hog->model(idx_multi_hog)->info()->cell_size();
167  const size_t num_bins = multi_hog->model(idx_multi_hog)->info()->num_bins();
168 
169  // Calculate number of cells along the x and y directions for the hog_space
170  const size_t num_cells_x = width / cell.width;
171  const size_t num_cells_y = height / cell.height;
172 
173  // TensorShape of hog space
174  TensorShape shape_hog_space = input->info()->tensor_shape();
175  shape_hog_space.set(Window::DimX, num_cells_x);
176  shape_hog_space.set(Window::DimY, num_cells_y);
177 
178  // Allocate HOG space
179  TensorInfo info_space(shape_hog_space, num_bins, DataType::F32);
180  _hog_space[i].allocator()->init(info_space);
181 
182  // Manage intermediate buffers
183  _memory_group.manage(&_hog_space[i]);
184 
185  // Initialise orientation binning kernel
186  _orient_bin_kernel[i].configure(&_mag, &_phase, &_hog_space[i], multi_hog->model(idx_multi_hog)->info());
187  }
188 
189  // Allocate intermediate tensors
190  _mag.allocator()->allocate();
191  _phase.allocator()->allocate();
192 
193  // Configure NETensor for the normalized HOG space and block normalization kernel
194  for(size_t i = 0; i < _num_block_norm_kernel; ++i)
195  {
196  const size_t idx_multi_hog = input_block_norm[i].first;
197  const size_t idx_orient_bin = input_block_norm[i].second;
198 
199  // Allocate normalized HOG space
200  TensorInfo tensor_info(*(multi_hog->model(idx_multi_hog)->info()), width, height);
201  _hog_norm_space[i].allocator()->init(tensor_info);
202 
203  // Manage intermediate buffers
204  _memory_group.manage(&_hog_norm_space[i]);
205 
206  // Initialize block normalization kernel
207  _block_norm_kernel[i].configure(&_hog_space[idx_orient_bin], &_hog_norm_space[i], multi_hog->model(idx_multi_hog)->info());
208  }
209 
210  // Allocate intermediate tensors
211  for(size_t i = 0; i < _num_orient_bin_kernel; ++i)
212  {
213  _hog_space[i].allocator()->allocate();
214  }
215 
216  // Configure HOG detector kernel
217  for(size_t i = 0; i < _num_hog_detect_kernel; ++i)
218  {
219  const size_t idx_block_norm = input_hog_detect[i];
220 
221  _hog_detect_kernel[i].configure(&_hog_norm_space[idx_block_norm], multi_hog->model(i), detection_windows, detection_window_strides->at(i), threshold, i);
222  }
223 
224  // Configure non maxima suppression kernel
225  _non_maxima_kernel.configure(_detection_windows, min_distance);
226 
227  // Allocate intermediate tensors
228  for(size_t i = 0; i < _num_block_norm_kernel; ++i)
229  {
230  _hog_norm_space[i].allocator()->allocate();
231  }
232 }
233 
235 {
236  ARM_COMPUTE_ERROR_ON_MSG(_detection_windows == nullptr, "Unconfigured function");
237 
238  MemoryGroupResourceScope scope_mg(_memory_group);
239 
240  // Reset detection window
241  _detection_windows->clear();
242 
243  // Run gradient
244  _gradient_kernel.run();
245 
246  // Run orientation binning kernel
247  for(auto &kernel : _orient_bin_kernel)
248  {
250  }
251 
252  // Run block normalization kernel
253  for(auto &kernel : _block_norm_kernel)
254  {
256  }
257 
258  // Run HOG detector kernel
259  for(auto &kernel : _hog_detect_kernel)
260  {
261  kernel.run();
262  }
263 
264  // Run non-maxima suppression kernel if enabled
265  if(_non_maxima_suppression)
266  {
267  NEScheduler::get().schedule(&_non_maxima_kernel, Window::DimY);
268  }
269 }
270 } // namespace arm_compute
BorderMode
Methods available to handle borders.
Definition: Types.h:265
size_t num_bins() const
The number of histogram bins for each cell.
Definition: HOGInfo.cpp:111
void clear()
Clear all the points from the array.
Definition: IArray.h:91
Shape of a tensor.
Definition: TensorShape.h:39
void init(const TensorAllocator &allocator, const Coordinates &coords, TensorInfo &sub_info)
Shares the same backing memory with another tensor allocator, while the tensor info might be differen...
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
const Size2D & cell_size() const
The cell size in pixels.
Definition: HOGInfo.cpp:91
PhaseType
Phase calculation type.
Definition: Types.h:432
1 channel, 1 U8 per channel
~NEHOGMultiDetection()
Default destructor.
Array of type T.
Definition: IArray.h:40
1 channel, 1 F32 per channel
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
PhaseType phase_type() const
The type of PhaseType.
Definition: HOGInfo.cpp:126
Interface for Neon tensor.
Definition: ITensor.h:36
const Size2D & block_stride() const
The block stride in pixels.
Definition: HOGInfo.cpp:106
Copyright (c) 2017-2021 Arm Limited.
size_t height
Height of the image region or rectangle.
Definition: Size2D.h:90
void run() override
Run the kernels contained in the function.
TensorAllocator * allocator()
Return a pointer to the tensor&#39;s allocator.
Definition: Tensor.cpp:48
void configure(ITensor *input, ITensor *output_magnitude, ITensor *output_phase, PhaseType phase_type, BorderMode border_mode, uint8_t constant_border_value=0)
Initialise the function&#39;s source, destinations, phase type and border mode.
virtual size_t num_models() const =0
The number of HOG models stored.
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
#define ARM_COMPUTE_ERROR_ON_INVALID_MULTI_HOG(m)
Definition: Validate.h:925
size_t num_values() const
Number of values currently stored in the array.
Definition: IArray.h:68
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
#define ARM_COMPUTE_ERROR_ON_MSG(cond, msg)
Definition: Error.h:456
const Size2D & block_size() const
The block size in pixels.
Definition: HOGInfo.cpp:96
void allocate() override
Allocate size specified by TensorInfo of CPU memory.
void configure(IDetectionWindowArray *input_output, float min_distance)
Initialise the kernel&#39;s input, output and the euclidean minimum distance.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
void run() override
Run the kernels contained in the function.
1 channel, 1 S16 per channel
#define ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:790
CPP kernel to perform in-place computation of euclidean distance on IDetectionWindowArray.
void configure(ITensor *input, const IMultiHOG *multi_hog, IDetectionWindowArray *detection_windows, const ISize2DArray *detection_window_strides, BorderMode border_mode, uint8_t constant_border_value=0, float threshold=0.0f, bool non_maxima_suppression=false, float min_distance=1.0f)
Initialise the function&#39;s source, destination, detection window strides, border mode, threshold and non-maxima suppression.
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
virtual IHOG * model(size_t index)=0
Return a pointer to the requested HOG model.
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
virtual void schedule(ICPPKernel *kernel, const Hints &hints)=0
Runs the kernel in the same thread as the caller synchronously.
size_t width
Width of the image region or rectangle.
Definition: Size2D.h:89
Interface for storing multiple HOG data-objects.
Definition: IMultiHOG.h:34
SimpleTensor< T > non_maxima_suppression(const SimpleTensor< T > &src, BorderMode border_mode, T constant_border_value)
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:45
NEHOGMultiDetection(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
virtual const HOGInfo * info() const =0
Interface to be implemented by the child class to return the HOG&#39;s metadata.
SimpleTensor< T > threshold(const SimpleTensor< T > &src, T threshold, T false_value, T true_value, ThresholdType type, T upper)
Definition: Threshold.cpp:35
TensorShape & set(size_t dimension, size_t value, bool apply_dim_correction=true, bool increase_dim_unit=true)
Accessor to set the value of one of the dimensions.
Definition: TensorShape.h:79
virtual T & at(size_t index) const
Reference to the element of the array located at the given index.
Definition: IArray.h:117
static IScheduler & get()
Access the scheduler singleton.
Definition: Scheduler.cpp:94