search

Megvii-BaseDetection / OTA

Official implementation of our CVPR2021 paper "OTA: Optimal Transport Assignment for Object Detection" in Pytorch.
Apache License 2.0
243 stars 24 forks source link

Reproduction #2

Closed strongwolf closed 3 years ago

strongwolf commented 3 years ago

I changed nothing for the config file but only got 39.9 mAP when training res50 by myself. And I tested with your provided weight file and only got 31 mAP for res50 and 36 for res101.

Joker316701882 commented 3 years ago

@strongwolf About the low mAP for our provided weight files, you are right, I'll work on it soon.

Joker316701882 commented 3 years ago

@strongwolf About the low 39.9mAP for your own training. Did you use the latest version of cvpods?

strongwolf commented 3 years ago

@strongwolf About the low 39.9mAP for your own training. Did you use the latest version of cvpods?

I used version 0.1.

Joker316701882 commented 3 years ago

@strongwolf Hi, I re-trained an ota.res50 and got the following result: image

So, I guess you need to update your cvpods to the latest version first, then check whether your COCO dataset is complete. Currently, I only upload the re-trained res50, the other models will come soon.