Yuwen Xiong

Please check our paper, one labeled image per snippet is enough for our framework in training phase.

If you use FPN, make sure your GPUs have at least 12G memory.

Please list your environment clearly, including CUDA version, Caffe version (please ensure you read the README and use the Caffe version we suggest); whether this situation is reproducible. And make...

You could just take res5c layer's output and flatten it, but please note that the vector length will not be a constant if your input image size varies.

Hi @Xiangyu-CAS and @mursalal , please try our new implementation, https://github.com/msracver/Deformable-ConvNets. It supports multi-GPU training/test and require GPU memory.

loss_bbox = 0 means all the boxes are back-ground boxes. Your data might be extremely unbalanced.