The results were not reproduced on CrowdHuman dataset

[a1b2c3s4d4]

The CroedHuman dataset is converted by "./SparseR-CNN-main/projects/SparseRCNN/crowdhumantools/convert_crowdhuman_to_coco.py". And the config file we used is "./SparseR-CNN-main/projects/SparseRCNN/configs/sparsercnn.crowdhuman.res50.500pro.50e.yaml". When training, we didnot change any training parameters in the code. The server we used is GeForce 3090 with 4 cards.

But the mAP we obtain is 79.325 on CrowdHuman val, which is much lower than the mAP reported in the paper. The mAP reported in the paper is 89.2.

Can you help to analyze why the results cannot be reproduced?

Thank you very much!

[PeizeSun]

Hi, Your evaluation result is by COCO tool, not official CrowdHuman tool. Please follow the evaluation step 4 here