search

megvii-research / Iter-E2EDET

Official implementation of the paper "Progressive End-to-End Object Detection in Crowded Scenes"
MIT License
88 stars 8 forks source link

How to obtain the same accuracy as the paper? #11

Closed ma3252788 closed 1 year ago

ma3252788 commented 2 years ago

Hello, may I ask why, according to your instructions, when I directly run this command python3 train_net.py --num-gpus 1 --config-file configs/50e.6h.500pro.ignore.yaml --eval-only MODEL.WEIGHTS ../../e2edet_final.pth, there is only AP value and no MR. in addition, the AP value is also very low, only 82%

Thank you very much for any reply

image

yexiguafuqihao commented 2 years ago

emm, please run demo.py in each project for evaluation.

ma3252788 commented 2 years ago

Thank you for your reply, but it seems that there is no demo.py file in the project

image

yexiguafuqihao commented 2 years ago

You can download the provided model for fast evaluation, whose performance is reported in our paper. Also, we use Pascal-VOC style evaluation code for $\text{AP}$ and $\text{MR}^{-2}$ evaluation, which is a bit different from that in the COCO. Sorry, the demo.py is renamed as eval_result.py when we shared the code.

image