can't reproduce the result present in the paper

Jeff-sjtu / res-loglikelihood-regression

Code for "Human Pose Regression with Residual Log-likelihood Estimation", ICCV 2021 Oral

423 stars 44 forks source link

can't reproduce the result present in the paper #38

Closed zimenglan-sysu-512 closed 2 years ago

zimenglan-sysu-512 commented 2 years ago

hi @Jeff-sjtu i use the cmd ./scripts/train.sh ./configs/256x192_res50_regress-flow.yaml train_rle_coco to reproduce the result present in the paper, but only get ##### Epoch 255 | gt mAP: 0.7123198166036238 | det mAP: 0.6962231312113063 #####, lower than the released model you provide where gt box: 0.7218652898214926 mAP | det box: 0.7127219006071578 mAP. can u give the complete training pipeline to reproduce the result? thanks

zimenglan-sysu-512 commented 2 years ago

btw, i use two GPUs to train.

zimenglan-sysu-512 commented 2 years ago

hi @Jeff-sjtu now i can reproduce the result present in the paper,

train a heatmap based model, which get 4 GPUs ==> Epoch 139/140 | gt mAP: 0.7253/0.7264/111 | det mAP: 0.7101/0.7101/121
use the trained model to finetune rlepose, and get Test Epoch 255/270 | gt mAP: 0.7257/0.7267/227 | det mAP: 0.7128/0.7131/219

markkim1115 commented 2 years ago

Hello, small curiosity, can you share what the problem was???

zimenglan-sysu-512 commented 2 years ago

hi @markkim1115 directly train rle, can't get the results from released model. so you should pretrain a heatmap-based model.

Jackqu commented 2 years ago

hi， which heatmap-based model does you use ? @zimenglan-sysu-512