search

AlbertoSabater / Robust-and-efficient-post-processing-for-video-object-detection

GNU General Public License v3.0
146 stars 20 forks source link

normalize output repp #19

Closed M-Hajakbari closed 2 years ago

M-Hajakbari commented 3 years ago

Hello,

Thanks for sharing this nice repo.

I would like normalize output repp for get map with yolo4 program. how i can do it? output repp xyxy or xywh?

thanks

sylivahf commented 3 years ago

hello, i also want to add the post-processing into yolov5's detect.py. but i got the poor predictions. did you have the better predictions?

AlbertoSabater commented 3 years ago

Hello. I am not familiar with YOLO 4 and 5. REPP uses the COCO format (xywh), which eases the later mAP calculation with pycocotools.

sylivahf commented 3 years ago

Hello. I am not familiar with YOLO 4 and 5. REPP uses the COCO format (xywh), which eases the later mAP calculation with pycocotools.

thank you for sharing!

yeah! i uses the coco format (like this: {1: [{'id': 0, 'bbox': [472, 110, 16, 20], 'bbox_center': [480, 120, 16, 20], 'scores': [ 0.37517, 0.0020256, 0.00066277, 0.0032884]}, {'id': 1, 'bbox': [582, 361, 16, 12], 'bbox_center': [590, 367, 16, 12], 'scores':[ 0.49756, 0.0009155, 0.00011327, 0.0051832], ....}. where key (1) is frame index, key-id is index of detected candidate target, key-scores is class confidence（there are 4 classes).

and, i only use two adjacent frames，the previous frame data is still original data (no repp operation). then, there are the following problems :

  1. missing detecttion is compared to NMS
  2. there are multiple overlapping boxes for the same target

Is there something wrong with me in understanding REPP?