How it Modify for YOLOV8 pose for multi classes #190
Comments
Sure you should find the top label by scores_ptr. |
|
Even if all objects were detected and output, the location could only be estimated for one class. The 17 location points are also wrong. Still, objects of other classes are not even detected. Objects are not detected properly before labeling. *. YoloV8-TensorRt Program result(8 object Detect, but almost same position and size) *. Ultralytics Cli result |
|
Do you have further questions? |
The model learned with multi-class(3) is performed normally with the yolov8 cli Predict command.
I confirmed through debugging that only one class objects was detected when using the example here.
*. you recommed.
You should modify pose postprocess.
Default is only one class so that the score is only for person.
But if enable multiclass you should use argmax and max to find the highest score and label.
=> Where should I fix it?
*. I check it.
YOLOv8-TensorRT\csrc\pose\normal\include\yolov8-pose.hpp
....
void YOLOv8_pose::postprocess(std::vector& objs, float score_thres, float iou_thres, int topk)
{
...
auto kps_ptr = row_ptr + 5;
// classnum = 3
auto maxSPtr = std::max_element(kps_ptr, kps_ptr + 3);
// auto maxSPtr = std::max_element(kps_ptr+2, kps_ptr + 3);
// auto maxSPtr = std::max_element(kps_ptr+3, kps_ptr + 3);
// Not Detect label
cout << func << ";label=" << label << endl;
.....
int cnt = 0;
/*
#ifdef BATCHED_NMS
cv::dnn::NMSBoxesBatched(bboxes, scores, labels, score_thresd, iou_thres, indices);
#else
cv::dnn::NMSBoxes(bboxes, scores, score_thres, iou_thres, indices);
#endif
for(auto& i : indices)
*/
// 8 object Detect, but almost same position and size.
for(int i = 0; i < bboxes.size(); i++)
{
if(cnt >= topk)
break;
Object obj;
obj.rect = bboxes[i];
obj.prob = scores[i];
obj.label = labels[i];
obj.kps = kpss[i];
objs.push_back(obj);
cnt += 1;
}
}
....
The text was updated successfully, but these errors were encountered: