search

Zhongdao / Towards-Realtime-MOT

Joint Detection and Embedding for fast multi-object tracking
MIT License
2.37k stars 539 forks source link

mAP 0 on custom dataset for Car tracking #255

Open DariusCorp opened 1 year ago

DariusCorp commented 1 year ago

Hello, I am using a custom 1000 images dataset like FLIR ADAS for car tracking. I have successfully converted all the data to fit the example datasets, and the model can train on it. The problem is that mAP is always 0 even after 300 epochs. I am running on Nvidia Quadro RTX 5000, batch size is 8, my image size is 640x480, the cfg chosen are 864x480. What could be the problem here? I tried running on different cfg but still no improvement

These are training losses for epoch 12: 2023-02-28 14:32:24 [INFO]: 12/29 60/86 0.0313 0.00521 2.2 -10.9 61 0.54

manish181192 commented 1 year ago

Dis you figure out whats wrong?