@@ -11,12 +11,16 @@ namespace YoloV5
1111 internal class Program
1212 {
1313
14- #if YOLOV5_V60
14+ #region Fields
15+
16+ #if YOLOV5_V60 || YOLOV5_V62
1517 private const int MaxStride = 64 ;
1618#else
1719 private const int MaxStride = 32 ;
1820#endif
1921
22+ #endregion
23+
2024 private sealed class YoloV5Focus : CustomLayer
2125 {
2226
@@ -38,12 +42,13 @@ protected override unsafe int OnForward(Mat bottomBlob, Mat topBlob, Option opt)
3842 topBlob . Create ( outW , outH , outC , 4u , 1 , opt . BlobAllocator ) ;
3943 if ( topBlob . IsEmpty )
4044 return - 100 ;
41-
45+
46+ // ToDo: parallel
4247 for ( var p = 0 ; p < outC ; p ++ )
4348 {
4449 using var bottom = bottomBlob . Channel ( p % channels ) ;
4550 using var top = topBlob . Channel ( p ) ;
46- var ptr = ( float * ) ( bottom . Row ( ( p / channels ) % 2 ) . Data + ( ( p / channels ) / 2 ) ) ;
51+ var ptr = ( float * ) ( bottom . Row ( ( p / channels ) % 2 ) . Data ) + ( ( p / channels ) / 2 ) ;
4752 var outPtr = ( float * ) top . Data ;
4853
4954 for ( var i = 0 ; i < outH ; i ++ )
@@ -90,14 +95,14 @@ private static int Main(string[] args)
9095 return - 1 ;
9196 }
9297
93- if ( Ncnn . IsSupportVulkan )
94- Ncnn . CreateGpuInstance ( ) ;
98+ // if (Ncnn.IsSupportVulkan)
99+ // Ncnn.CreateGpuInstance();
95100
96101 var objects = new List < Object > ( ) ;
97102 DetectYoloV5 ( m , objects ) ;
98103
99- if ( Ncnn . IsSupportVulkan )
100- Ncnn . DestroyGpuInstance ( ) ;
104+ // if (Ncnn.IsSupportVulkan)
105+ // Ncnn.DestroyGpuInstance();
101106
102107 DrawObjects ( m , objects ) ;
103108 }
@@ -223,61 +228,61 @@ private static void GenerateProposals(Mat anchors, int stride, Mat inPad, Mat fe
223228 for ( var j = 0 ; j < numGridX ; j ++ )
224229 {
225230 var featPtr = feat . Row ( i * numGridX + j ) ;
226-
227- // find class index with max class score
228- var classIndex = 0 ;
229- var classScore = - float . MaxValue ;
230- for ( var k = 0 ; k < numClass ; k ++ )
231+ var boxConfidence = Sigmoid ( featPtr [ 4 ] ) ;
232+ if ( boxConfidence >= probThreshold )
231233 {
232- var score = featPtr [ 5 + k ] ;
233- if ( score > classScore )
234+ // find class index with max class score
235+ var classIndex = 0 ;
236+ var classScore = - float . MaxValue ;
237+ for ( var k = 0 ; k < numClass ; k ++ )
234238 {
235- classIndex = k ;
236- classScore = score ;
239+ var score = featPtr [ 5 + k ] ;
240+ if ( score > classScore )
241+ {
242+ classIndex = k ;
243+ classScore = score ;
244+ }
237245 }
238- }
239-
240- var box_score = featPtr [ 4 ] ;
241246
242- var confidence = Sigmoid ( box_score ) * Sigmoid ( classScore ) ;
243-
244- if ( confidence >= probThreshold )
245- {
246- // yolov5/models/yolo.py Detect forward
247- // y = x[i].Sigmoid()
248- // y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy
249- // y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
247+ var confidence = boxConfidence * Sigmoid ( classScore ) ;
248+ if ( confidence >= probThreshold )
249+ {
250+ // yolov5/models/yolo.py Detect forward
251+ // y = x[i].Sigmoid()
252+ // y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy
253+ // y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
250254
251- var dx = Sigmoid ( featPtr [ 0 ] ) ;
252- var dy = Sigmoid ( featPtr [ 1 ] ) ;
253- var dw = Sigmoid ( featPtr [ 2 ] ) ;
254- var dh = Sigmoid ( featPtr [ 3 ] ) ;
255+ var dx = Sigmoid ( featPtr [ 0 ] ) ;
256+ var dy = Sigmoid ( featPtr [ 1 ] ) ;
257+ var dw = Sigmoid ( featPtr [ 2 ] ) ;
258+ var dh = Sigmoid ( featPtr [ 3 ] ) ;
255259
256- var pbCx = ( dx * 2.0f - 0.5f + j ) * stride ;
257- var pbCy = ( dy * 2.0f - 0.5f + i ) * stride ;
260+ var pbCx = ( dx * 2.0f - 0.5f + j ) * stride ;
261+ var pbCy = ( dy * 2.0f - 0.5f + i ) * stride ;
258262
259- var pbW = ( float ) Math . Pow ( dw * 2.0f , 2 ) * anchorW ;
260- var pbH = ( float ) Math . Pow ( dh * 2.0f , 2 ) * anchorH ;
263+ var pbW = ( float ) Math . Pow ( dw * 2.0f , 2 ) * anchorW ;
264+ var pbH = ( float ) Math . Pow ( dh * 2.0f , 2 ) * anchorH ;
261265
262- var x0 = pbCx - pbW * 0.5f ;
263- var y0 = pbCy - pbH * 0.5f ;
264- var x1 = pbCx + pbW * 0.5f ;
265- var y1 = pbCy + pbH * 0.5f ;
266+ var x0 = pbCx - pbW * 0.5f ;
267+ var y0 = pbCy - pbH * 0.5f ;
268+ var x1 = pbCx + pbW * 0.5f ;
269+ var y1 = pbCy + pbH * 0.5f ;
266270
267- var obj = new Object
268- {
269- Rect =
271+ var obj = new Object
272+ {
273+ Rect =
270274 {
271275 X = x0 ,
272276 Y = y0 ,
273277 Width = x1 - x0 ,
274278 Height = y1 - y0
275279 } ,
276- Label = classIndex ,
277- Prob = confidence
278- } ;
280+ Label = classIndex ,
281+ Prob = confidence
282+ } ;
279283
280- objects . Add ( obj ) ;
284+ objects . Add ( obj ) ;
285+ }
281286 }
282287 }
283288 }
@@ -294,7 +299,10 @@ private static void DetectYoloV5(NcnnDotNet.OpenCV.Mat bgr, List<Object> objects
294299
295300 // original pretrained model from https://github.com/ultralytics/yolov5
296301 // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models
297- #if YOLOV5_V60
302+ #if YOLOV5_V62
303+ yolov5 . LoadParam ( "yolov5s_6.2.param" ) ;
304+ yolov5 . LoadModel ( "yolov5s_6.2.bin" ) ;
305+ #elif YOLOV5_V60
298306 yolov5 . LoadParam ( "yolov5s_6.0.param" ) ;
299307 yolov5 . LoadModel ( "yolov5s_6.0.bin" ) ;
300308#else
@@ -371,7 +379,9 @@ private static void DetectYoloV5(NcnnDotNet.OpenCV.Mat bgr, List<Object> objects
371379 // stride 16
372380 {
373381 using var @out = new Mat ( ) ;
374- #if YOLOV5_V60
382+ #if YOLOV5_V62
383+ ex . Extract ( "353" , @out ) ;
384+ #elif YOLOV5_V60
375385 ex . Extract ( "376" , @out ) ;
376386#else
377387 ex . Extract ( "781" , @out ) ;
@@ -394,7 +404,9 @@ private static void DetectYoloV5(NcnnDotNet.OpenCV.Mat bgr, List<Object> objects
394404 // stride 32
395405 {
396406 using var @out = new Mat ( ) ;
397- #if YOLOV5_V60
407+ #if YOLOV5_V62
408+ ex . Extract ( "367" , @out ) ;
409+ #elif YOLOV5_V60
398410 ex . Extract ( "401" , @out ) ;
399411#else
400412 ex . Extract ( "801" , @out ) ;
0 commit comments