Open May-forever opened 5 years ago
@hjm1990818 Hi,
All mAP and speed are taken from referenced artices (authors of DPM, RCNN, ...)
https://en.wikipedia.org/wiki/F1_score that you see as F1-score
when run ./darknet detector map ...
I sort all detections by Confidence-score, then go through all objects and take into account only 1 detection, then 2 detections, 3 detections.... to build Precision-Recall curve, then take 11 points on it. https://medium.com/@jonathan_hui/map-mean-average-precision-for-object-detection-45c121a31173
You can look at this code: https://github.com/AlexeyAB/darknet/blob/5343aa423563c107e0071b1427ad5defc27b56d2/src/detector.c#L638-L1000
Dear AlexeyAB, @AlexeyAB Sorry to bother you, but I really need your help.
1. In Figure 5(a) of paper 'You only look once: unified, real-time object detection', you compared YOLO with DPM,Poselets and RCNN. However, I did not find the .cfgs and weights of DPM,Poselets and RCNN in darknet-master, could please tell me where to download them?
2. The second question is what's the meaning of 'Best F1' ? Please see Figure 5(b).
3. How do you get the 11 points of the PR value? By changing a special threshold 11 times ?
I am looking forward to hearing from you, thank you very much in advance.
With best regards