RFCN_TensorFlow

Results

model type	training strategy	mAP(%) on VOC07 test	Iterations	model_name	backbone
conv5, a trous, strides=16 without ohem	4 stages iteration as Faster RCNN	75.77	total steps 400k satge1 80k stage2 120k stage3 80k stage4 120k	model_A	resnet_101
conv5, a trous, strides=16 without ohem	only training total_loss	76.35	110k	model_B	resnet_101

total_loss = loss_rpn_objectness + loss_rpn_bboxes + loss_rfcn_classes + loss_rfcn_bboxes

Result Details

model_name	aeroplane	bicycle	bird	boat	bottle	bus	car	cat	chair	cow	diningtable	dog	horse	motorbike	person	pottedplant	sheep	sofa	train	tvmonitor
model_A	0.8008	0.8004	0.7861	0.6579	0.4836	0.8646	0.8531	0.8774	0.6081	0.8517	0.6935	0.8884	0.8616	0.7821	0.7805	0.4693	0.7814	0.7742	0.7845	0.7516
model_B	0.8020	0.7940	0.7877	0.6402	0.6571	0.8599	0.8578	0.8736	0.6183	0.8223	0.6492	0.8728	0.8447	0.8201	0.7888	0.4607	0.7703	0.7558	0.8354	0.7596

Training Details

model_A

  momentum: 0.9
  stage1 total steps 80k, init learning rate 0.001, step 60k learning rate 0.0001
  stage2 total steps 120k, init learning rate 0.001, step 80k learning rate 0.0001
  stage3 total steps 80k, init learning rate 0.001, step 60k learning rate 0.0001
  stage4 total steps 120k, init learning rate 0.001, step 80k learning rate 0.0001

model_B

  momentum: 0.9
  total steps 110k, init learning rate 0.001, step 80k learning rate 0.0001

Model Download Links

model_name	download link	password
model_A	https://pan.baidu.com/s/1jIQThtW	cgwf
model_B	https://pan.baidu.com/s/1i4QEVRZ	v9ua

Tasks

• ohem (I have tried several methods, but have no effect. The map in all the methods have dropped.)
• focal loss (The focal loss also have no effect.)
• position sensitive score map + global roi pooling class.
• code refactor

Training Pipline

1. running tools/trainval_net_rfcn.py file.
   • modify the net you want to use in line import nets, the nets provied by this project is in floder lib/nets.
   • modify the --cfg parameter which is the config file you want to use. Some config file can be find in file experiments/cfgs.
   • modify the --weight parameter which is the pretrained model weights file.
   • modify the --net which is the net architecture you want to use.
2. some other modifies:
   • you can modify the loss function as you requirement in line network.py#L646
   • you can modify the RFCN architecture in line resnet_v1_rfcn_hole.py#L279_
   • you'd better using the resnet_v1_rfcn_hole.py and network.py file.

References:

1 tf-faster-rcnn

2 tensorflow-object-detection

3 R-FCN: Object Detection via Region-based Fully Convolutional Networks

4 An Implementation of Faster RCNN with Study for Region Sampling