cannot get good training result

[FCInter]

I try to train the model from the very beginning.

I downloaded the models and put them as following

./model/pretrained_model/resnet_v1_101-0000.params
./model/pretrained_model/flownet-0000.params

Then I started the training using the default configuration. But the model cannot detect any object. I have tested after training 1, 2, ..., 8 epoches. The training loss looks good. I got RPNAcc about 0.98, and RCNNAcc about 0.90. But the test mAP was merely 10 or 20, and in the demo, no object was detected.

By comparison, if I trained from the checkpoint you provided, i.e., the following:

./model/rfcn_vid-0000.params
./model/rfcn_dff_flownet_vid-0000.params

After 2 epoches, I got even better test and demo results than the provided checkpoints. Both on rfcn and rfcn_dff_flownet.

The only difference between the two experiments, is whether loading the provided checkpoint when training.

I was wondering why this happens.

BTW, I also run the code of FGFA, which is also built by your group. That code was good. Even if I trained from the very beginning (only loaded flownet and resnet, didn't load the provided checkpoint), I got better test results after 3 or 4 epoches.

Why can't I get good results this time?

Thank you!

[xuwentang]

Try contacting the author.I contacted the author, but he did not respond.

[BoyuanFeng]

"Then I started the training using the default configuration. But the model cannot detect any object. I have tested after training 1, 2, ..., 8 epoches. The training loss looks good. I got RPNAcc about 0.98, and RCNNAcc about 0.90. But the test mAP was merely 10 or 20, and in the demo, no object was detected"

I confronted a similar issue.