Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
John Richardson | 8de9261 | 2018-02-22 14:09:31 +0000 | [diff] [blame] | 2 | * Copyright (c) 2016-2018 ARM Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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 | */ |
| 24 | #include "arm_compute/core/NEON/kernels/NELKTrackerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/Coordinates.h" |
| 28 | #include "arm_compute/core/Error.h" |
| 29 | #include "arm_compute/core/Helpers.h" |
| 30 | #include "arm_compute/core/ITensor.h" |
| 31 | #include "arm_compute/core/TensorInfo.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | #include "arm_compute/core/Window.h" |
| 34 | |
| 35 | #include <arm_neon.h> |
| 36 | #include <cmath> |
| 37 | |
| 38 | using namespace arm_compute; |
| 39 | |
| 40 | /** Constants used for Lucas-Kanade Algorithm */ |
| 41 | constexpr int W_BITS = 14; |
| 42 | constexpr float D0 = 1 << W_BITS; |
| 43 | constexpr float DETERMINANT_THRESHOLD = 1.0e-07f; // Threshold for the determinant. Used for lost tracking criteria |
| 44 | constexpr float EIGENVALUE_THRESHOLD = 1.0e-04f; // Thresholds for minimum eigenvalue. Used for lost tracking criteria |
| 45 | constexpr float FLT_SCALE = 1.0f / (1 << 20); |
| 46 | |
| 47 | namespace |
| 48 | { |
| 49 | enum class BilinearInterpolation |
| 50 | { |
| 51 | BILINEAR_OLD_NEW, |
| 52 | BILINEAR_SCHARR |
| 53 | }; |
| 54 | |
| 55 | template <typename T> |
| 56 | constexpr int INT_ROUND(T x, int n) |
| 57 | { |
| 58 | return (x + (1 << (n - 1))) >> n; |
| 59 | } |
| 60 | |
| 61 | template <typename T> |
| 62 | inline int get_pixel(const ITensor *tensor, int xi, int yi, int iw00, int iw01, int iw10, int iw11, int scale) |
| 63 | { |
| 64 | const auto px00 = *reinterpret_cast<const T *>(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi, yi))); |
| 65 | const auto px01 = *reinterpret_cast<const T *>(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi + 1, yi))); |
| 66 | const auto px10 = *reinterpret_cast<const T *>(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi, yi + 1))); |
| 67 | const auto px11 = *reinterpret_cast<const T *>(tensor->buffer() + tensor->info()->offset_element_in_bytes(Coordinates(xi + 1, yi + 1))); |
| 68 | |
| 69 | return INT_ROUND(px00 * iw00 + px01 * iw01 + px10 * iw10 + px11 * iw11, scale); |
| 70 | } |
| 71 | |
| 72 | inline int32x4_t compute_bilinear_interpolation(int16x8_t top_row, int16x8_t bottom_row, int16x4_t w00, int16x4_t w01, int16x4_t w10, int16x4_t w11, int32x4_t shift) |
| 73 | { |
| 74 | // Get the left column of upper row |
| 75 | const int16x4_t px00 = vget_low_s16(top_row); |
| 76 | |
| 77 | // Get the right column of upper row |
| 78 | const int16x4_t px01 = vext_s16(px00, vget_high_s16(top_row), 1); |
| 79 | |
| 80 | // Get the left column of lower row |
| 81 | const int16x4_t px10 = vget_low_s16(bottom_row); |
| 82 | |
| 83 | // Get the right column of right row |
| 84 | const int16x4_t px11 = vext_s16(px10, vget_high_s16(bottom_row), 1); |
| 85 | |
| 86 | // Apply the bilinear filter |
| 87 | return vqrshlq_s32(vmull_s16(px00, w00) + vmull_s16(px01, w01) + vmull_s16(px10, w10) + vmull_s16(px11, w11), shift); |
| 88 | } |
| 89 | } // namespace |
| 90 | |
| 91 | void NELKTrackerKernel::init_keypoints(int start, int end) |
| 92 | { |
| 93 | if(_level == _num_levels - 1) |
| 94 | { |
| 95 | const float level_scale = pow(_pyramid_scale, _level); |
| 96 | |
| 97 | for(int i = start; i < end; ++i) |
| 98 | { |
| 99 | _old_points_internal->at(i).x = _old_points->at(i).x * level_scale; |
| 100 | _old_points_internal->at(i).y = _old_points->at(i).y * level_scale; |
| 101 | _old_points_internal->at(i).tracking_status = true; |
| 102 | |
| 103 | NELKInternalKeypoint keypoint_to_track; |
| 104 | |
| 105 | if(_use_initial_estimate) |
| 106 | { |
| 107 | keypoint_to_track.x = _new_points_estimates->at(i).x * level_scale; |
| 108 | keypoint_to_track.y = _new_points_estimates->at(i).y * level_scale; |
| 109 | keypoint_to_track.tracking_status = (_new_points_estimates->at(i).tracking_status == 1); |
| 110 | } |
| 111 | else |
| 112 | { |
| 113 | keypoint_to_track.x = _old_points_internal->at(i).x; |
| 114 | keypoint_to_track.y = _old_points_internal->at(i).y; |
| 115 | keypoint_to_track.tracking_status = true; |
| 116 | } |
| 117 | |
| 118 | _new_points_internal->at(i) = keypoint_to_track; |
| 119 | } |
| 120 | } |
| 121 | else |
| 122 | { |
| 123 | for(int i = start; i < end; ++i) |
| 124 | { |
| 125 | _old_points_internal->at(i).x /= _pyramid_scale; |
| 126 | _old_points_internal->at(i).y /= _pyramid_scale; |
| 127 | _new_points_internal->at(i).x /= _pyramid_scale; |
| 128 | _new_points_internal->at(i).y /= _pyramid_scale; |
| 129 | } |
| 130 | } |
| 131 | } |
| 132 | |
Michalis Spyrou | 490bf2e | 2017-09-29 11:24:55 +0100 | [diff] [blame] | 133 | std::tuple<int, int, int> NELKTrackerKernel::compute_spatial_gradient_matrix(const NELKInternalKeypoint &keypoint, int32_t *bilinear_ix, int32_t *bilinear_iy) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 134 | { |
| 135 | int iA11 = 0; |
| 136 | int iA12 = 0; |
| 137 | int iA22 = 0; |
| 138 | |
| 139 | int32x4_t nA11 = vdupq_n_s32(0); |
| 140 | int32x4_t nA12 = vdupq_n_s32(0); |
| 141 | int32x4_t nA22 = vdupq_n_s32(0); |
| 142 | |
| 143 | float keypoint_int_x = 0; |
| 144 | float keypoint_int_y = 0; |
| 145 | |
| 146 | const float wx = std::modf(keypoint.x, &keypoint_int_x); |
| 147 | const float wy = std::modf(keypoint.y, &keypoint_int_y); |
| 148 | |
| 149 | const int iw00 = roundf((1.0f - wx) * (1.0f - wy) * D0); |
| 150 | const int iw01 = roundf(wx * (1.0f - wy) * D0); |
| 151 | const int iw10 = roundf((1.0f - wx) * wy * D0); |
| 152 | const int iw11 = D0 - iw00 - iw01 - iw10; |
| 153 | |
| 154 | const int16x4_t nw00 = vdup_n_s16(iw00); |
| 155 | const int16x4_t nw01 = vdup_n_s16(iw01); |
| 156 | const int16x4_t nw10 = vdup_n_s16(iw10); |
| 157 | const int16x4_t nw11 = vdup_n_s16(iw11); |
| 158 | |
| 159 | // Convert stride from uint_t* to int16_t* |
| 160 | const size_t row_stride = _old_scharr_gx->info()->strides_in_bytes()[1] / 2; |
| 161 | const Coordinates top_left_window_corner(static_cast<int>(keypoint_int_x) - _window_dimension / 2, static_cast<int>(keypoint_int_y) - _window_dimension / 2); |
| 162 | auto idx = reinterpret_cast<const int16_t *>(_old_scharr_gx->buffer() + _old_scharr_gx->info()->offset_element_in_bytes(top_left_window_corner)); |
| 163 | auto idy = reinterpret_cast<const int16_t *>(_old_scharr_gy->buffer() + _old_scharr_gy->info()->offset_element_in_bytes(top_left_window_corner)); |
| 164 | static const int32x4_t nshifter_scharr = vdupq_n_s32(-W_BITS); |
| 165 | |
| 166 | for(int ky = 0; ky < _window_dimension; ++ky, idx += row_stride, idy += row_stride) |
| 167 | { |
| 168 | int kx = 0; |
| 169 | |
| 170 | // Calculate elements in blocks of four as long as possible |
| 171 | for(; kx <= _window_dimension - 4; kx += 4) |
| 172 | { |
| 173 | // Interpolation X |
| 174 | const int16x8_t ndx_row1 = vld1q_s16(idx + kx); |
| 175 | const int16x8_t ndx_row2 = vld1q_s16(idx + kx + row_stride); |
| 176 | |
| 177 | const int32x4_t nxval = compute_bilinear_interpolation(ndx_row1, ndx_row2, nw00, nw01, nw10, nw11, nshifter_scharr); |
| 178 | |
| 179 | // Interpolation Y |
| 180 | const int16x8_t ndy_row1 = vld1q_s16(idy + kx); |
| 181 | const int16x8_t ndy_row2 = vld1q_s16(idy + kx + row_stride); |
| 182 | |
| 183 | const int32x4_t nyval = compute_bilinear_interpolation(ndy_row1, ndy_row2, nw00, nw01, nw10, nw11, nshifter_scharr); |
| 184 | |
| 185 | // Store the intermediate data so that we don't need to recalculate them in later stage |
| 186 | vst1q_s32(bilinear_ix + kx + ky * _window_dimension, nxval); |
| 187 | vst1q_s32(bilinear_iy + kx + ky * _window_dimension, nyval); |
| 188 | |
| 189 | // Accumulate Ix^2 |
| 190 | nA11 = vmlaq_s32(nA11, nxval, nxval); |
| 191 | // Accumulate Ix * Iy |
| 192 | nA12 = vmlaq_s32(nA12, nxval, nyval); |
| 193 | // Accumulate Iy^2 |
| 194 | nA22 = vmlaq_s32(nA22, nyval, nyval); |
| 195 | } |
| 196 | |
| 197 | // Calculate the leftover elements |
| 198 | for(; kx < _window_dimension; ++kx) |
| 199 | { |
| 200 | const int32_t ixval = get_pixel<int16_t>(_old_scharr_gx, top_left_window_corner.x() + kx, top_left_window_corner.y() + ky, |
| 201 | iw00, iw01, iw10, iw11, W_BITS); |
| 202 | const int32_t iyval = get_pixel<int16_t>(_old_scharr_gy, top_left_window_corner.x() + kx, top_left_window_corner.y() + ky, |
| 203 | iw00, iw01, iw10, iw11, W_BITS); |
| 204 | |
| 205 | iA11 += ixval * ixval; |
| 206 | iA12 += ixval * iyval; |
| 207 | iA22 += iyval * iyval; |
| 208 | |
| 209 | bilinear_ix[kx + ky * _window_dimension] = ixval; |
| 210 | bilinear_iy[kx + ky * _window_dimension] = iyval; |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | iA11 += vgetq_lane_s32(nA11, 0) + vgetq_lane_s32(nA11, 1) + vgetq_lane_s32(nA11, 2) + vgetq_lane_s32(nA11, 3); |
| 215 | iA12 += vgetq_lane_s32(nA12, 0) + vgetq_lane_s32(nA12, 1) + vgetq_lane_s32(nA12, 2) + vgetq_lane_s32(nA12, 3); |
| 216 | iA22 += vgetq_lane_s32(nA22, 0) + vgetq_lane_s32(nA22, 1) + vgetq_lane_s32(nA22, 2) + vgetq_lane_s32(nA22, 3); |
| 217 | |
| 218 | return std::make_tuple(iA11, iA12, iA22); |
| 219 | } |
| 220 | |
Michalis Spyrou | 490bf2e | 2017-09-29 11:24:55 +0100 | [diff] [blame] | 221 | std::pair<int, int> NELKTrackerKernel::compute_image_mismatch_vector(const NELKInternalKeypoint &old_keypoint, const NELKInternalKeypoint &new_keypoint, const int32_t *bilinear_ix, |
| 222 | const int32_t *bilinear_iy) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 223 | { |
| 224 | int ib1 = 0; |
| 225 | int ib2 = 0; |
| 226 | |
| 227 | int32x4_t nb1 = vdupq_n_s32(0); |
| 228 | int32x4_t nb2 = vdupq_n_s32(0); |
| 229 | |
| 230 | // Compute weights for the old keypoint |
| 231 | float old_keypoint_int_x = 0; |
| 232 | float old_keypoint_int_y = 0; |
| 233 | |
| 234 | const float old_wx = std::modf(old_keypoint.x, &old_keypoint_int_x); |
| 235 | const float old_wy = std::modf(old_keypoint.y, &old_keypoint_int_y); |
| 236 | |
| 237 | const int iw00_old = roundf((1.0f - old_wx) * (1.0f - old_wy) * D0); |
| 238 | const int iw01_old = roundf(old_wx * (1.0f - old_wy) * D0); |
| 239 | const int iw10_old = roundf((1.0f - old_wx) * old_wy * D0); |
| 240 | const int iw11_old = D0 - iw00_old - iw01_old - iw10_old; |
| 241 | |
| 242 | const int16x4_t nw00_old = vdup_n_s16(iw00_old); |
| 243 | const int16x4_t nw01_old = vdup_n_s16(iw01_old); |
| 244 | const int16x4_t nw10_old = vdup_n_s16(iw10_old); |
| 245 | const int16x4_t nw11_old = vdup_n_s16(iw11_old); |
| 246 | |
| 247 | // Compute weights for the new keypoint |
| 248 | float new_keypoint_int_x = 0; |
| 249 | float new_keypoint_int_y = 0; |
| 250 | |
| 251 | const float new_wx = std::modf(new_keypoint.x, &new_keypoint_int_x); |
| 252 | const float new_wy = std::modf(new_keypoint.y, &new_keypoint_int_y); |
| 253 | |
| 254 | const int iw00_new = roundf((1.0f - new_wx) * (1.0f - new_wy) * D0); |
| 255 | const int iw01_new = roundf(new_wx * (1.0f - new_wy) * D0); |
| 256 | const int iw10_new = roundf((1.0f - new_wx) * new_wy * D0); |
| 257 | const int iw11_new = D0 - iw00_new - iw01_new - iw10_new; |
| 258 | |
| 259 | const int16x4_t nw00_new = vdup_n_s16(iw00_new); |
| 260 | const int16x4_t nw01_new = vdup_n_s16(iw01_new); |
| 261 | const int16x4_t nw10_new = vdup_n_s16(iw10_new); |
| 262 | const int16x4_t nw11_new = vdup_n_s16(iw11_new); |
| 263 | |
| 264 | const int row_stride = _input_new->info()->strides_in_bytes()[1]; |
| 265 | const Coordinates top_left_window_corner_old(static_cast<int>(old_keypoint_int_x) - _window_dimension / 2, static_cast<int>(old_keypoint_int_y) - _window_dimension / 2); |
| 266 | const Coordinates top_left_window_corner_new(static_cast<int>(new_keypoint_int_x) - _window_dimension / 2, static_cast<int>(new_keypoint_int_y) - _window_dimension / 2); |
| 267 | const uint8_t *old_ptr = _input_old->buffer() + _input_old->info()->offset_element_in_bytes(top_left_window_corner_old); |
| 268 | const uint8_t *new_ptr = _input_new->buffer() + _input_new->info()->offset_element_in_bytes(top_left_window_corner_new); |
| 269 | static const int32x4_t nshifter_tensor = vdupq_n_s32(-(W_BITS - 5)); |
| 270 | |
| 271 | for(int ky = 0; ky < _window_dimension; ++ky, new_ptr += row_stride, old_ptr += row_stride) |
| 272 | { |
| 273 | int kx = 0; |
| 274 | |
| 275 | // Calculate elements in blocks of four as long as possible |
| 276 | for(; kx <= _window_dimension - 4; kx += 4) |
| 277 | { |
| 278 | // Interpolation old tensor |
| 279 | const int16x8_t nold_row1 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(old_ptr + kx))); |
| 280 | const int16x8_t nold_row2 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(old_ptr + kx + row_stride))); |
| 281 | |
| 282 | const int32x4_t noldval = compute_bilinear_interpolation(nold_row1, nold_row2, nw00_old, nw01_old, nw10_old, nw11_old, nshifter_tensor); |
| 283 | |
| 284 | // Interpolation new tensor |
| 285 | const int16x8_t nnew_row1 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(new_ptr + kx))); |
| 286 | const int16x8_t nnew_row2 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(new_ptr + kx + row_stride))); |
| 287 | |
| 288 | const int32x4_t nnewval = compute_bilinear_interpolation(nnew_row1, nnew_row2, nw00_new, nw01_new, nw10_new, nw11_new, nshifter_tensor); |
| 289 | |
| 290 | // Calculate It gradient, i.e. pixelwise difference between old and new tensor |
| 291 | const int32x4_t diff = vsubq_s32(nnewval, noldval); |
| 292 | |
| 293 | // Load the Ix and Iy gradient computed in the previous stage |
| 294 | const int32x4_t nxval = vld1q_s32(bilinear_ix + kx + ky * _window_dimension); |
| 295 | const int32x4_t nyval = vld1q_s32(bilinear_iy + kx + ky * _window_dimension); |
| 296 | |
| 297 | // Caculate Ix * It and Iy * It, and accumulate the results |
| 298 | nb1 = vmlaq_s32(nb1, diff, nxval); |
| 299 | nb2 = vmlaq_s32(nb2, diff, nyval); |
| 300 | } |
| 301 | |
| 302 | // Calculate the leftover elements |
| 303 | for(; kx < _window_dimension; ++kx) |
| 304 | { |
| 305 | const int32_t ival = get_pixel<uint8_t>(_input_old, top_left_window_corner_old.x() + kx, top_left_window_corner_old.y() + ky, |
| 306 | iw00_old, iw01_old, iw10_old, iw11_old, W_BITS - 5); |
| 307 | const int32_t jval = get_pixel<uint8_t>(_input_new, top_left_window_corner_new.x() + kx, top_left_window_corner_new.y() + ky, |
| 308 | iw00_new, iw01_new, iw10_new, iw11_new, W_BITS - 5); |
| 309 | |
| 310 | const int32_t diff = jval - ival; |
| 311 | |
| 312 | ib1 += diff * bilinear_ix[kx + ky * _window_dimension]; |
| 313 | ib2 += diff * bilinear_iy[kx + ky * _window_dimension]; |
| 314 | } |
| 315 | } |
| 316 | |
| 317 | ib1 += vgetq_lane_s32(nb1, 0) + vgetq_lane_s32(nb1, 1) + vgetq_lane_s32(nb1, 2) + vgetq_lane_s32(nb1, 3); |
| 318 | ib2 += vgetq_lane_s32(nb2, 0) + vgetq_lane_s32(nb2, 1) + vgetq_lane_s32(nb2, 2) + vgetq_lane_s32(nb2, 3); |
| 319 | |
| 320 | return std::make_pair(ib1, ib2); |
| 321 | } |
| 322 | |
| 323 | NELKTrackerKernel::NELKTrackerKernel() |
| 324 | : _input_old(nullptr), _input_new(nullptr), _old_scharr_gx(nullptr), _old_scharr_gy(nullptr), _new_points(nullptr), _new_points_estimates(nullptr), _old_points(nullptr), _old_points_internal(), |
| 325 | _new_points_internal(), _termination(Termination::TERM_CRITERIA_EPSILON), _use_initial_estimate(false), _pyramid_scale(0.0f), _epsilon(0.0f), _num_iterations(0), _window_dimension(0), _level(0), |
| 326 | _num_levels(0), _valid_region() |
| 327 | { |
| 328 | } |
| 329 | |
| 330 | BorderSize NELKTrackerKernel::border_size() const |
| 331 | { |
| 332 | return BorderSize(1); |
| 333 | } |
| 334 | |
| 335 | void NELKTrackerKernel::configure(const ITensor *input_old, const ITensor *input_new, const ITensor *old_scharr_gx, const ITensor *old_scharr_gy, |
| 336 | const IKeyPointArray *old_points, const IKeyPointArray *new_points_estimates, IKeyPointArray *new_points, |
| 337 | INELKInternalKeypointArray *old_points_internal, INELKInternalKeypointArray *new_points_internal, |
| 338 | Termination termination, bool use_initial_estimate, float epsilon, unsigned int num_iterations, size_t window_dimension, |
| 339 | size_t level, size_t num_levels, float pyramid_scale) |
| 340 | |
| 341 | { |
| 342 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_old, 1, DataType::U8); |
| 343 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_new, 1, DataType::U8); |
| 344 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(old_scharr_gx, 1, DataType::S16); |
| 345 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(old_scharr_gy, 1, DataType::S16); |
| 346 | |
| 347 | _input_old = input_old; |
| 348 | _input_new = input_new; |
| 349 | _old_scharr_gx = old_scharr_gx; |
| 350 | _old_scharr_gy = old_scharr_gy; |
| 351 | _old_points = old_points; |
| 352 | _new_points_estimates = new_points_estimates; |
| 353 | _new_points = new_points; |
| 354 | _old_points_internal = old_points_internal; |
| 355 | _new_points_internal = new_points_internal; |
| 356 | _termination = termination; |
| 357 | _use_initial_estimate = use_initial_estimate; |
| 358 | _epsilon = epsilon; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 359 | _window_dimension = window_dimension; |
| 360 | _level = level; |
| 361 | _num_levels = num_levels; |
| 362 | _pyramid_scale = pyramid_scale; |
| 363 | _num_levels = num_levels; |
| 364 | |
John Richardson | 8de9261 | 2018-02-22 14:09:31 +0000 | [diff] [blame] | 365 | // Set maximum number of iterations used for convergence |
| 366 | const size_t max_iterations = 1000; |
| 367 | _num_iterations = (termination == Termination::TERM_CRITERIA_EPSILON) ? max_iterations : num_iterations; |
| 368 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 369 | Window window; |
| 370 | window.set(Window::DimX, Window::Dimension(0, old_points->num_values())); |
| 371 | window.set(Window::DimY, Window::Dimension(0, 1)); |
| 372 | |
| 373 | _valid_region = intersect_valid_regions( |
| 374 | input_old->info()->valid_region(), |
| 375 | input_new->info()->valid_region(), |
| 376 | old_scharr_gx->info()->valid_region(), |
| 377 | old_scharr_gy->info()->valid_region()); |
| 378 | |
| 379 | update_window_and_padding(window, |
| 380 | AccessWindowStatic(input_old->info(), _valid_region.start(0), _valid_region.start(1), |
| 381 | _valid_region.end(0), _valid_region.end(1)), |
| 382 | AccessWindowStatic(input_new->info(), _valid_region.start(0), _valid_region.start(1), |
| 383 | _valid_region.end(0), _valid_region.end(1)), |
| 384 | AccessWindowStatic(old_scharr_gx->info(), _valid_region.start(0), _valid_region.start(1), |
| 385 | _valid_region.end(0), _valid_region.end(1)), |
| 386 | AccessWindowStatic(old_scharr_gy->info(), _valid_region.start(0), _valid_region.start(1), |
| 387 | _valid_region.end(0), _valid_region.end(1))); |
| 388 | |
| 389 | INEKernel::configure(window); |
| 390 | } |
| 391 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 392 | void NELKTrackerKernel::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 393 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 394 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 395 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 396 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 397 | |
| 398 | ARM_COMPUTE_ERROR_ON(_input_old->buffer() == nullptr); |
| 399 | ARM_COMPUTE_ERROR_ON(_input_new->buffer() == nullptr); |
| 400 | ARM_COMPUTE_ERROR_ON(_old_scharr_gx->buffer() == nullptr); |
| 401 | ARM_COMPUTE_ERROR_ON(_old_scharr_gy->buffer() == nullptr); |
| 402 | |
| 403 | const int list_end = window.x().end(); |
| 404 | const int list_start = window.x().start(); |
| 405 | |
| 406 | init_keypoints(list_start, list_end); |
| 407 | |
| 408 | const int buffer_size = _window_dimension * _window_dimension; |
Michalis Spyrou | 490bf2e | 2017-09-29 11:24:55 +0100 | [diff] [blame] | 409 | int32_t bilinear_ix[buffer_size]; |
| 410 | int32_t bilinear_iy[buffer_size]; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 411 | |
| 412 | const int half_window = _window_dimension / 2; |
| 413 | |
| 414 | auto is_invalid_keypoint = [&](const NELKInternalKeypoint & keypoint) |
| 415 | { |
| 416 | const int x = std::floor(keypoint.x); |
| 417 | const int y = std::floor(keypoint.y); |
| 418 | |
| 419 | return (x - half_window < _valid_region.start(0)) || (x + half_window >= _valid_region.end(0) - 1) || (y - half_window < _valid_region.start(1)) || (y + half_window >= _valid_region.end(1) - 1); |
| 420 | }; |
| 421 | |
| 422 | for(int list_indx = list_start; list_indx < list_end; ++list_indx) |
| 423 | { |
| 424 | NELKInternalKeypoint &old_keypoint = _old_points_internal->at(list_indx); |
| 425 | NELKInternalKeypoint &new_keypoint = _new_points_internal->at(list_indx); |
| 426 | |
| 427 | if(!old_keypoint.tracking_status) |
| 428 | { |
| 429 | continue; |
| 430 | } |
| 431 | |
| 432 | if(is_invalid_keypoint(old_keypoint)) |
| 433 | { |
| 434 | if(_level == 0) |
| 435 | { |
| 436 | new_keypoint.tracking_status = false; |
| 437 | } |
| 438 | |
| 439 | continue; |
| 440 | } |
| 441 | |
| 442 | // Compute spatial gradient matrix |
| 443 | int iA11 = 0; |
| 444 | int iA12 = 0; |
| 445 | int iA22 = 0; |
| 446 | |
| 447 | std::tie(iA11, iA12, iA22) = compute_spatial_gradient_matrix(old_keypoint, bilinear_ix, bilinear_iy); |
| 448 | |
| 449 | const float A11 = iA11 * FLT_SCALE; |
| 450 | const float A12 = iA12 * FLT_SCALE; |
| 451 | const float A22 = iA22 * FLT_SCALE; |
| 452 | |
| 453 | // Calculate minimum eigenvalue |
| 454 | const float sum_A11_A22 = A11 + A22; |
| 455 | const float discriminant = sum_A11_A22 * sum_A11_A22 - 4.0f * (A11 * A22 - A12 * A12); |
| 456 | // Divide by _window_dimension^2 to reduce the floating point accummulation error |
| 457 | const float minimum_eigenvalue = (sum_A11_A22 - std::sqrt(discriminant)) / (2.0f * _window_dimension * _window_dimension); |
| 458 | |
| 459 | // Determinant |
| 460 | const double D = A11 * A22 - A12 * A12; |
| 461 | |
| 462 | // Check if it is a good point to track |
| 463 | if(minimum_eigenvalue < EIGENVALUE_THRESHOLD || D < DETERMINANT_THRESHOLD) |
| 464 | { |
| 465 | // Invalidate tracked point |
| 466 | if(_level == 0) |
| 467 | { |
| 468 | new_keypoint.tracking_status = false; |
| 469 | } |
| 470 | |
| 471 | continue; |
| 472 | } |
| 473 | |
| 474 | float prev_delta_x = 0.0f; |
| 475 | float prev_delta_y = 0.0f; |
| 476 | |
John Richardson | 8de9261 | 2018-02-22 14:09:31 +0000 | [diff] [blame] | 477 | for(unsigned int j = 0; j < _num_iterations; ++j) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 478 | { |
| 479 | if(is_invalid_keypoint(new_keypoint)) |
| 480 | { |
| 481 | if(_level == 0) |
| 482 | { |
| 483 | new_keypoint.tracking_status = false; |
| 484 | } |
| 485 | |
| 486 | break; |
| 487 | } |
| 488 | |
| 489 | // Compute image mismatch vector |
| 490 | int ib1 = 0; |
| 491 | int ib2 = 0; |
| 492 | |
| 493 | std::tie(ib1, ib2) = compute_image_mismatch_vector(old_keypoint, new_keypoint, bilinear_ix, bilinear_iy); |
| 494 | |
| 495 | double b1 = ib1 * FLT_SCALE; |
| 496 | double b2 = ib2 * FLT_SCALE; |
| 497 | |
| 498 | // Compute motion vector -> A^-1 * -b |
| 499 | const float delta_x = (A12 * b2 - A22 * b1) / D; |
| 500 | const float delta_y = (A12 * b1 - A11 * b2) / D; |
| 501 | |
| 502 | // Update the new position |
| 503 | new_keypoint.x += delta_x; |
| 504 | new_keypoint.y += delta_y; |
| 505 | |
| 506 | const float mag2 = delta_x * delta_x + delta_y * delta_y; |
| 507 | |
| 508 | // Check if termination criteria is EPSILON and if it is satisfied |
| 509 | if(mag2 <= _epsilon && (_termination == Termination::TERM_CRITERIA_EPSILON || _termination == Termination::TERM_CRITERIA_BOTH)) |
| 510 | { |
| 511 | break; |
| 512 | } |
| 513 | |
| 514 | // Check convergence analyzing the previous delta |
| 515 | if(j > 0 && std::fabs(delta_x + prev_delta_x) < 0.01f && std::fabs(delta_y + prev_delta_y) < 0.01f) |
| 516 | { |
| 517 | new_keypoint.x -= delta_x * _pyramid_scale; |
| 518 | new_keypoint.y -= delta_y * _pyramid_scale; |
| 519 | break; |
| 520 | } |
| 521 | |
| 522 | prev_delta_x = delta_x; |
| 523 | prev_delta_y = delta_y; |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | if(_level == 0) |
| 528 | { |
| 529 | for(int list_indx = list_start; list_indx < list_end; ++list_indx) |
| 530 | { |
| 531 | const NELKInternalKeypoint &new_keypoint = _new_points_internal->at(list_indx); |
| 532 | |
| 533 | _new_points->at(list_indx).x = roundf(new_keypoint.x); |
| 534 | _new_points->at(list_indx).y = roundf(new_keypoint.y); |
| 535 | _new_points->at(list_indx).tracking_status = new_keypoint.tracking_status ? 1 : 0; |
| 536 | } |
| 537 | } |
| 538 | } |