ligang-cs/csp-pedestrian-detection

A faster pytorch implementation of 'Center and Scale Prediction (CSP) for pedestrain detection (CVPR19)'.

• ✨ I am Gang Li (李钢), a researcher at Tencent Youtu Lab. I am particularly interested in Large Language Model alignments and agents.
• 🌱 Before joining Tencent, I receieved my Ph.D degree from Nanjing Univeristy of Science and Technology in 2023, under the supervision of Prof. Xiang Li and Prof. Shanshan Zhang.
• 👯 During my PhD study, I spent wonderful research time at SenseTime Research (mentored by Yujie Wang and Ding Liang) and Shanghai AI Lab (mentored by Wenhai Wang). I also interned at ByteDance and Tencent.

csp-pedestrian-detection's Issues

How to reproduce the performance in README?

@ligang-cs Hi, great work!
However, I wondered how to reproduce the reported 11.3 MR performance in README (I mean, how to modify the config.py)? I ran you code and only got 13.8 for reasonable (for Res50). And also, when I use 8 GPUs to train, should I do some extra modifications on the config.py?

Loss is NaN

Great work！It is easy to understand and modify codes.
But when i try to train CSP（resnet50 as backbone）， I met a Problem：Loss is NaN. Even i low lr, it also happen.
So,I want to ask for some suggestions for this problem