tf-faster-rcnn-relation-module

A reimplementation of relation module for faster rcnn in Tensorflow (refer to paper Relation Network for Object Detection). This repository is based on the multi-image version of Faster RCNN available here. The codes for relation module are based on the original mxnet implementation.

./experiments/scripts/test_faster_rcnn.sh $GPU_ID pascal_voc_0712 res101 RM.ENABLE_INSTANCE True TRAIN.WARMUP True

As above shell shown, we set RM.ENABLE_INSTANCE as True to enable relation module. The learning rate is set according to the original mxnet implementation. More configuration for relation module can be viewed in lib/model/config.py

Note:

• Only the relation module (between fc layers for feature enhance) is implemented for present. The one for duplicate process (to learn NMS) in not support yet.
• relation module is impletemented in lib/nets/network.py.
• The position embedding code are implemented in python, and reside in lib/layer_utils/rel_module_util.py.

Update(0717):

• Add warm up support, refer to lib/model/config.py for more details.
• Add ZFnet for quick iteration, the model is converted from caffe model, and the pool layers are changed from kernel 3, stride 2 to kernel 2, stride 2