Light_head_R_CNN_xception

The source code in this repository contain Light-Head R-CNN: In Defense of Two-Stage Object Detector that use xception* backbone network. It is based on the code from zengarden. I tried it with Pascal VOC and COCO dataset. The test mAP and FPS of the original paper has not yet been achieved. It would be great if someone would give me a advice. The network architecture is below from original paper.

Requirement:

Same as in here.

Installation:

1. Clone the Light_head_R_CNN_xception repository.

git clone https://github.com/geonseoks/Light_head_R_CNN_xception

2. To use the xception* networks, the ckpt files for xception* has to be downloaded. Move the ckpt file into ~your_light_head_rcnn_original_directory/data/imagenet_weights.

3. Move the resnet_v1.py and resnet_utils.py into ~your_light_head_rcnn_original_directory/lib/utils/tf_utils/basemodel.

4. Move the network_desp.py into ~your_light_head_rcnn_original_directory/experiments/user/network_desp.py.

File structure as follows:

~Your_light_head_rcnn_original_directory/
    |->data
    |    |->imagenet_weights
    |    |    |->model.ckpt-3081378.ckpt
    |->experiments
    |    |->user
    |    |    |->network_desp.py
    |->lib
    |    |->utils
    |    |    |->tf_utils
    |    |    |    |->basemodel
    |    |    |    |    |->resnet_v1.py
    |    |    |    |    |->resnet_utils.py    

Train:

Same as in here except learning rate (basic_lr = 5e-4 * train_batch_per_gpu * 0.7 is better).

Evalution:

Same as in here.

Result:

The bold text on the top line is the result of the original paper.

Train data	Test data	ImageNet backbone accuracy at 224*224 (%)	Base model	Input resolution	GPU	FPS	Epochs	mAP (%)
MSCOCO	MSCOCO	65.9	xception*	700x1100	TITAN XP	102	-	30.7
MSCOCO	MSCOCO	65.0	xception*	700x1100	GTX 1080Ti	51.89	30	26.1
MSCOCO	MSCOCO	65.0	xception*	700x1100	TITAN X PASCAL	31.0	30	26.1
VOC07	VOC07	65.0	xception*	700x1100	GTX 1080Ti	54.07	-	62.0
VOC07	VOC07	65.0	xception*	700x1100	TITAN X PASCAL	33.4	-	62.0
VOC07	VOC07	65.0	xception*	144x144	TITAN X PASCAL	164.0	-	56.9
VOC07+VOC12	VOC07	65.0	xception*	144x144	TITAN X PASCAL	174.0	-	61.0
VOC07	VOC07	65.0	xception*	128x128	TITAN X PASCAL	180.4	-	55.0
VOC07+VOC12	VOC07	65.0	xception*	128x128	TITAN X PASCAL	-	-	59.9