• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * By downloading, copying, installing or using the software you agree to this license.
3  * If you do not agree to this license, do not download, install,
4  * copy or use the software.
5  *
6  *
7  *                           License Agreement
8  *                For Open Source Computer Vision Library
9  *                        (3-clause BSD License)
10  *
11  * Copyright (C) 2012-2015, NVIDIA Corporation, all rights reserved.
12  * Third party copyrights are property of their respective owners.
13  *
14  * Redistribution and use in source and binary forms, with or without modification,
15  * are permitted provided that the following conditions are met:
16  *
17  *   * Redistributions of source code must retain the above copyright notice,
18  *     this list of conditions and the following disclaimer.
19  *
20  *   * Redistributions in binary form must reproduce the above copyright notice,
21  *     this list of conditions and the following disclaimer in the documentation
22  *     and/or other materials provided with the distribution.
23  *
24  *   * Neither the names of the copyright holders nor the names of the contributors
25  *     may be used to endorse or promote products derived from this software
26  *     without specific prior written permission.
27  *
28  * This software is provided by the copyright holders and contributors "as is" and
29  * any express or implied warranties, including, but not limited to, the implied
30  * warranties of merchantability and fitness for a particular purpose are disclaimed.
31  * In no event shall copyright holders or contributors be liable for any direct,
32  * indirect, incidental, special, exemplary, or consequential damages
33  * (including, but not limited to, procurement of substitute goods or services;
34  * loss of use, data, or profits; or business interruption) however caused
35  * and on any theory of liability, whether in contract, strict liability,
36  * or tort (including negligence or otherwise) arising in any way out of
37  * the use of this software, even if advised of the possibility of such damage.
38  */
39 
40 
41 /* This is FAST corner detector, contributed to OpenCV by the author, Edward Rosten.
42    Below is the original copyright and the references */
43 
44 /*
45 Copyright (c) 2006, 2008 Edward Rosten
46 All rights reserved.
47 
48 Redistribution and use in source and binary forms, with or without
49 modification, are permitted provided that the following conditions
50 are met:
51 
52  *Redistributions of source code must retain the above copyright
53   notice, this list of conditions and the following disclaimer.
54 
55  *Redistributions in binary form must reproduce the above copyright
56   notice, this list of conditions and the following disclaimer in the
57   documentation and/or other materials provided with the distribution.
58 
59  *Neither the name of the University of Cambridge nor the names of
60   its contributors may be used to endorse or promote products derived
61   from this software without specific prior written permission.
62 
63 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
64 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
65 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
66 A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
67 CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
68 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
69 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
70 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
71 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
72 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
73 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
74 */
75 
76 /*
77 The references are:
78  * Machine learning for high-speed corner detection,
79    E. Rosten and T. Drummond, ECCV 2006
80  * Faster and better: A machine learning approach to corner detection
81    E. Rosten, R. Porter and T. Drummond, PAMI, 2009
82 */
83 
84 #include "common.hpp"
85 
86 #include <vector>
87 #include <cstring>
88 
89 namespace CAROTENE_NS {
90 
91 #ifdef CAROTENE_NEON
92 namespace
93 {
94 
makeOffsets(ptrdiff_t pixel[],ptrdiff_t row_stride)95 void makeOffsets(ptrdiff_t pixel[], ptrdiff_t row_stride)
96 {
97     pixel[0] = 0 + row_stride * 3;
98     pixel[1] = 1 + row_stride * 3;
99     pixel[2] = 2 + row_stride * 2;
100     pixel[3] = 3 + row_stride * 1;
101     pixel[4] = 3 + row_stride * 0;
102     pixel[5] = 3 + row_stride * -1;
103     pixel[6] = 2 + row_stride * -2;
104     pixel[7] = 1 + row_stride * -3;
105     pixel[8] = 0 + row_stride * -3;
106     pixel[9] = -1 + row_stride * -3;
107     pixel[10] = -2 + row_stride * -2;
108     pixel[11] = -3 + row_stride * -1;
109     pixel[12] = -3 + row_stride * 0;
110     pixel[13] = -3 + row_stride * 1;
111     pixel[14] = -2 + row_stride * 2;
112     pixel[15] = -1 + row_stride * 3;
113 }
114 
cornerScore(const u8 * ptr,const ptrdiff_t pixel[])115 u8 cornerScore(const u8* ptr, const ptrdiff_t pixel[])
116 {
117     const s32 K = 8, N = 16 + K + 1;
118     s32 k, v = ptr[0];
119     s16 d[(N + 7) & ~7];
120     for( k = 0; k < N; k++ )
121         d[k] = (s16)(v - ptr[pixel[k]]);
122 
123     int16x8_t q0 = vdupq_n_s16((s16)(-1000));
124     int16x8_t q1 = vdupq_n_s16((s16)(1000));
125 
126     int16x8_t d0_7   = vld1q_s16(d +  0);
127     int16x8_t d8_15  = vld1q_s16(d +  8);
128     int16x8_t d16_23 = vld1q_s16(d + 16);
129     int16x8_t d24    = vld1q_s16(d + 24);
130 
131     //k == 0
132     int16x8_t v0k0 = vextq_s16(d0_7, d8_15, 1);
133     int16x8_t v1k0 = vextq_s16(d0_7, d8_15, 2);
134     int16x8_t ak0 = vminq_s16(v0k0, v1k0);
135     int16x8_t bk0 = vmaxq_s16(v0k0, v1k0);
136 
137     v0k0 = vextq_s16(d0_7, d8_15, 3);
138     ak0 = vminq_s16(ak0, v0k0);
139     bk0 = vmaxq_s16(bk0, v0k0);
140 
141     v1k0 = vextq_s16(d0_7, d8_15, 4);
142     ak0 = vminq_s16(ak0, v1k0);
143     bk0 = vmaxq_s16(bk0, v1k0);
144 
145     v0k0 = vextq_s16(d0_7, d8_15, 5);
146     ak0 = vminq_s16(ak0, v0k0);
147     bk0 = vmaxq_s16(bk0, v0k0);
148 
149     v1k0 = vextq_s16(d0_7, d8_15, 6);
150     ak0 = vminq_s16(ak0, v1k0);
151     bk0 = vmaxq_s16(bk0, v1k0);
152 
153     v0k0 = vextq_s16(d0_7, d8_15, 7);
154     ak0 = vminq_s16(ak0, v0k0);
155     bk0 = vmaxq_s16(bk0, v0k0);
156 
157     ak0 = vminq_s16(ak0, d8_15);
158     bk0 = vmaxq_s16(bk0, d8_15);
159 
160     q0 = vmaxq_s16(q0, vminq_s16(ak0, d0_7));
161     q1 = vminq_s16(q1, vmaxq_s16(bk0, d0_7));
162 
163     v1k0 = vextq_s16(d8_15, d16_23, 1);
164     q0 = vmaxq_s16(q0, vminq_s16(ak0, v1k0));
165     q1 = vminq_s16(q1, vmaxq_s16(bk0, v1k0));
166 
167     //k == 8
168     int16x8_t v0k8 = v1k0;
169     int16x8_t v1k8 = vextq_s16(d8_15, d16_23, 2);
170     int16x8_t ak8 = vminq_s16(v0k8, v1k8);
171     int16x8_t bk8 = vmaxq_s16(v0k8, v1k8);
172 
173     v0k8 = vextq_s16(d8_15, d16_23, 3);
174     ak8 = vminq_s16(ak8, v0k8);
175     bk8 = vmaxq_s16(bk8, v0k8);
176 
177     v1k8 = vextq_s16(d8_15, d16_23, 4);
178     ak8 = vminq_s16(ak8, v1k8);
179     bk8 = vmaxq_s16(bk8, v1k8);
180 
181     v0k8 = vextq_s16(d8_15, d16_23, 5);
182     ak8 = vminq_s16(ak8, v0k8);
183     bk8 = vmaxq_s16(bk8, v0k8);
184 
185     v1k8 = vextq_s16(d8_15, d16_23, 6);
186     ak8 = vminq_s16(ak8, v1k8);
187     bk8 = vmaxq_s16(bk8, v1k8);
188 
189     v0k8 = vextq_s16(d8_15, d16_23, 7);
190     ak8 = vminq_s16(ak8, v0k8);
191     bk8 = vmaxq_s16(bk8, v0k8);
192 
193     ak8 = vminq_s16(ak8, d16_23);
194     bk8 = vmaxq_s16(bk8, d16_23);
195 
196     q0 = vmaxq_s16(q0, vminq_s16(ak8, d8_15));
197     q1 = vminq_s16(q1, vmaxq_s16(bk8, d8_15));
198 
199     v1k8 = vextq_s16(d16_23, d24, 1);
200     q0 = vmaxq_s16(q0, vminq_s16(ak8, v1k8));
201     q1 = vminq_s16(q1, vmaxq_s16(bk8, v1k8));
202 
203     //fin
204     int16x8_t q = vmaxq_s16(q0, vsubq_s16(vmovq_n_s16(0), q1));
205     int16x4_t q2 = vmax_s16(vget_low_s16(q), vget_high_s16(q));
206     int32x4_t q2w = vmovl_s16(q2);
207     int32x2_t q4 = vmax_s32(vget_low_s32(q2w), vget_high_s32(q2w));
208     int32x2_t q8 = vmax_s32(q4, vreinterpret_s32_s64(vshr_n_s64(vreinterpret_s64_s32(q4), 32)));
209 
210     return (u8)(vget_lane_s32(q8, 0) - 1);
211 }
212 
213 } //namespace
214 #endif
215 
FAST(const Size2D & size,u8 * srcBase,ptrdiff_t srcStride,KeypointStore * keypoints,u8 threshold,bool nonmax_suppression)216 void FAST(const Size2D &size,
217           u8 *srcBase, ptrdiff_t srcStride,
218           KeypointStore *keypoints,
219           u8 threshold, bool nonmax_suppression)
220 {
221     internal::assertSupportedConfiguration();
222 #ifdef CAROTENE_NEON
223     //keypoints.clear();
224 
225     const s32 K = 8, N = 16 + K + 1;
226     ptrdiff_t i, j, k, pixel[N];
227     makeOffsets(pixel, srcStride);
228     for(k = 16; k < N; k++)
229         pixel[k] = pixel[k - 16];
230 
231     uint8x16_t delta = vdupq_n_u8(128);
232     uint8x16_t t = vdupq_n_u8(threshold);
233     uint8x16_t K16 = vdupq_n_u8((u8)K);
234 
235     u8 threshold_tab[512];
236     for( i = -255; i <= 255; i++ )
237         threshold_tab[i+255] = (u8)(i < -threshold ? 1 : i > threshold ? 2 : 0);
238 
239     std::vector<u8> _buf((size.width+16)*3*(sizeof(ptrdiff_t) + sizeof(u8)) + 128);
240     u8* buf[3];
241     buf[0] = &_buf[0]; buf[1] = buf[0] + size.width; buf[2] = buf[1] + size.width;
242     ptrdiff_t* cpbuf[3];
243     cpbuf[0] = (ptrdiff_t*)internal::alignPtr(buf[2] + size.width, sizeof(ptrdiff_t)) + 1;
244     cpbuf[1] = cpbuf[0] + size.width + 1;
245     cpbuf[2] = cpbuf[1] + size.width + 1;
246     memset(buf[0], 0, size.width*3);
247 
248     for(i = 3; i < (ptrdiff_t)size.height-2; i++)
249     {
250         const u8* ptr = internal::getRowPtr(srcBase, srcStride, i) + 3;
251         u8* curr = buf[(i - 3)%3];
252         ptrdiff_t* cornerpos = cpbuf[(i - 3)%3];
253         memset(curr, 0, size.width);
254         ptrdiff_t ncorners = 0;
255 
256         if( i < (ptrdiff_t)size.height - 3 )
257         {
258             j = 3;
259 
260             for(; j < (ptrdiff_t)size.width - 16 - 3; j += 16, ptr += 16)
261             {
262                 internal::prefetch(ptr);
263                 internal::prefetch(ptr + pixel[0]);
264                 internal::prefetch(ptr + pixel[2]);
265 
266                 uint8x16_t v0 = vld1q_u8(ptr);
267                 int8x16_t v1 = vreinterpretq_s8_u8(veorq_u8(vqsubq_u8(v0, t), delta));
268                 int8x16_t v2 = vreinterpretq_s8_u8(veorq_u8(vqaddq_u8(v0, t), delta));
269 
270                 int8x16_t x0 = vreinterpretq_s8_u8(vsubq_u8(vld1q_u8(ptr + pixel[0]), delta));
271                 int8x16_t x1 = vreinterpretq_s8_u8(vsubq_u8(vld1q_u8(ptr + pixel[4]), delta));
272                 int8x16_t x2 = vreinterpretq_s8_u8(vsubq_u8(vld1q_u8(ptr + pixel[8]), delta));
273                 int8x16_t x3 = vreinterpretq_s8_u8(vsubq_u8(vld1q_u8(ptr + pixel[12]), delta));
274 
275                 uint8x16_t m0 =   vandq_u8(vcgtq_s8(x0, v2), vcgtq_s8(x1, v2));
276                 uint8x16_t m1 =   vandq_u8(vcgtq_s8(v1, x0), vcgtq_s8(v1, x1));
277                 m0 = vorrq_u8(m0, vandq_u8(vcgtq_s8(x1, v2), vcgtq_s8(x2, v2)));
278                 m1 = vorrq_u8(m1, vandq_u8(vcgtq_s8(v1, x1), vcgtq_s8(v1, x2)));
279                 m0 = vorrq_u8(m0, vandq_u8(vcgtq_s8(x2, v2), vcgtq_s8(x3, v2)));
280                 m1 = vorrq_u8(m1, vandq_u8(vcgtq_s8(v1, x2), vcgtq_s8(v1, x3)));
281                 m0 = vorrq_u8(m0, vandq_u8(vcgtq_s8(x3, v2), vcgtq_s8(x0, v2)));
282                 m1 = vorrq_u8(m1, vandq_u8(vcgtq_s8(v1, x3), vcgtq_s8(v1, x0)));
283                 m0 = vorrq_u8(m0, m1);
284 
285                 u64 mask[2];
286                 vst1q_u64(mask, vreinterpretq_u64_u8(m0));
287 
288                 if( mask[0] == 0 )
289                 {
290                     if (mask[1] != 0)
291                     {
292                         j -= 8;
293                         ptr -= 8;
294                     }
295                     continue;
296                 }
297 
298                 uint8x16_t c0 = vmovq_n_u8(0);
299                 uint8x16_t c1 = vmovq_n_u8(0);
300                 uint8x16_t max0 = vmovq_n_u8(0);
301                 uint8x16_t max1 = vmovq_n_u8(0);
302                 for( k = 0; k < N; k++ )
303                 {
304                     int8x16_t x = vreinterpretq_s8_u8(veorq_u8(vld1q_u8(ptr + pixel[k]), delta));
305                     m0 = vcgtq_s8(x, v2);
306                     m1 = vcgtq_s8(v1, x);
307 
308                     c0 = vandq_u8(vsubq_u8(c0, m0), m0);
309                     c1 = vandq_u8(vsubq_u8(c1, m1), m1);
310 
311                     max0 = vmaxq_u8(max0, c0);
312                     max1 = vmaxq_u8(max1, c1);
313                 }
314 
315                 max0 = vmaxq_u8(max0, max1);
316                 u8 m[16];
317                 vst1q_u8(m, vcgtq_u8(max0, K16));
318 
319                 for( k = 0; k < 16; ++k )
320                     if(m[k])
321                     {
322                         cornerpos[ncorners++] = j+k;
323                         if(nonmax_suppression)
324                             curr[j+k] = cornerScore(ptr+k, pixel);
325                     }
326             }
327 
328             for( ; j < (s32)size.width - 3; j++, ptr++ )
329             {
330                 s32 v = ptr[0];
331                 const u8* tab = &threshold_tab[0] - v + 255;
332                 s32 d = tab[ptr[pixel[0]]] | tab[ptr[pixel[8]]];
333 
334                 if( d == 0 )
335                     continue;
336 
337                 d &= tab[ptr[pixel[2]]] | tab[ptr[pixel[10]]];
338                 d &= tab[ptr[pixel[4]]] | tab[ptr[pixel[12]]];
339                 d &= tab[ptr[pixel[6]]] | tab[ptr[pixel[14]]];
340 
341                 if( d == 0 )
342                     continue;
343 
344                 d &= tab[ptr[pixel[1]]] | tab[ptr[pixel[9]]];
345                 d &= tab[ptr[pixel[3]]] | tab[ptr[pixel[11]]];
346                 d &= tab[ptr[pixel[5]]] | tab[ptr[pixel[13]]];
347                 d &= tab[ptr[pixel[7]]] | tab[ptr[pixel[15]]];
348 
349                 if( d & 1 )
350                 {
351                     s32 vt = v - threshold, count = 0;
352 
353                     for( k = 0; k < N; k++ )
354                     {
355                         s32 x = ptr[pixel[k]];
356                         if(x < vt)
357                         {
358                             if( ++count > K )
359                             {
360                                 cornerpos[ncorners++] = j;
361                                 if(nonmax_suppression)
362                                     curr[j] = cornerScore(ptr, pixel);
363                                 break;
364                             }
365                         }
366                         else
367                             count = 0;
368                     }
369                 }
370 
371                 if( d & 2 )
372                 {
373                     s32 vt = v + threshold, count = 0;
374 
375                     for( k = 0; k < N; k++ )
376                     {
377                         s32 x = ptr[pixel[k]];
378                         if(x > vt)
379                         {
380                             if( ++count > K )
381                             {
382                                 cornerpos[ncorners++] = j;
383                                 if(nonmax_suppression)
384                                     curr[j] = cornerScore(ptr, pixel);
385                                 break;
386                             }
387                         }
388                         else
389                             count = 0;
390                     }
391                 }
392             }
393         }
394 
395         cornerpos[-1] = ncorners;
396 
397         if( i == 3 )
398             continue;
399 
400         const u8* prev = buf[(i - 4 + 3)%3];
401         const u8* pprev = buf[(i - 5 + 3)%3];
402         cornerpos = cpbuf[(i - 4 + 3)%3];
403         ncorners = cornerpos[-1];
404 
405         for( k = 0; k < ncorners; k++ )
406         {
407             j = cornerpos[k];
408             s32 score = prev[j];
409             if( !nonmax_suppression ||
410                     (score > prev[j+1] && score > prev[j-1] &&
411                      score > pprev[j-1] && score > pprev[j] && score > pprev[j+1] &&
412                      score > curr[j-1] && score > curr[j] && score > curr[j+1]) )
413             {
414                 keypoints->push((f32)j, (f32)(i-1), 7.f, -1, (f32)score);
415             }
416         }
417     }
418 #else
419     (void)size;
420     (void)srcBase;
421     (void)srcStride;
422     (void)keypoints;
423     (void)threshold;
424     (void)nonmax_suppression;
425 #endif
426 }
427 
428 } // namespace CAROTENE_NS
429