Got poor performance on custom dataset.

I trained custom pascal_voc dataset with cascade_rcnn_hrnetv2p_w32_20e.py, I used all the data as trainval, and part of it as test, after 20 epochs, the loss decrease to 0.3, and the log info like this:

INFO - Epoch [16][450/715]    lr: 0.00250, eta: 3:52:14, time: 0.801, data_time: 0.004, memory: 9102, loss_rpn_cls: 0.0239, loss_rpn_bbox: 0.0093, s0.loss_cls: 0.1392, s0.acc: 96.1270, s0.loss_bbox: 0.0731, s1.loss_cls: 0.0549, s1.acc: 97.1621, s1.loss_bbox: 0.0480, s2.loss_cls: 0.0183, s2.acc: 98.4102, s2.loss_bbox: 0.0135, loss: 0.3802

And I used the checkpoint to run test.py and generate result.pkl, but after the I parsed the .pkl file, set score threshold 0.5, the result is not good, lots of images have no predictions. So what can I do to improve the performance or is there somewhere I forgot to modified. Thanks~

[wen0618]

@herrickli Have you solved your problem?My result is poor too. The loss is less than 0.0015 but the mAP is not good.

[xhqiui]

Maybe the pretrained in the configs isn’t suitable,you can try the default configuration in the section of the pretrained.