Training ImageNet and PASCAL VOC2012 via Learning Feature Pyramids

The code is provided by Guangrun Wang (Rongcong Chen also provides contributions).

Sun Yat-sen University (SYSU)

Table of Contents

0. Introduction
1. ImageNet
2. PASCAL VOC2012
3. Citation

Introduction

This repository contains the training & testing code on ImageNet and PASCAL VOC2012 via learning feature pyramids (LFP). LFP is originally used for human pose machine, described in the paper "Learning Feature Pyramids for Human Pose Estimation" (https://arxiv.org/abs/1708.01101). We extend it to the semantic image segmentation.

Results

• Segmentation Visualization:

  1. (a) input images; (b) segmentation results.

  2. (a) images & ground truths; (b) trimap of learning feature pyramids; (c) trimap of the original ResNet.

  3. It achieves 81.0% mIoU on PASCAL VOC2011 segmentation leaderboard, a significance improvement over its baseline DeepLabV2 (79.6%).

ImageNet

• Training script:

cd pyramid/ImageNet/
python imagenet-resnet.py   --gpu 0,1,2,3,4,5,6,7   --data_format NHWC  -d 101  --mode resnet --data  [ROOT-OF-IMAGENET-DATASET]

• Testing script:

cd pyramid/ImageNet/
python imagenet-resnet.py   --gpu 0,1,2,3,4,5,6,7  --load [ROOT-TO-LOAD-MODEL]  --data_format NHWC  -d 101  --mode resnet --data  [ROOT-OF-IMAGENET-DATASET] --eval

• Trained Models:

  ResNet101:

  Baidu Pan, code: 269o

  Google Drive

  ResNet50:

  Baidu Pan, code: zvgd

  Google Drive

PASCAL VOC2012

• Training script:

# Use the ImageNet classification model as pretrained model.
# Because ImageNet has 1,000 categories while voc only has 21 categories, 
# we must first fix all the parameters except the last layer including 21 channels. We only train the last layer for adaption
# by adding: "with freeze_variables(stop_gradient=True, skip_collection=True): " in Line 206 of resnet_model_voc_aspp.py
# Then we finetune all the parameters.
# For evaluation on voc val set, the model is first trained on COCO, then on train_aug of voc. 
# For evaluation on voc leaderboard (test set), the above model is further trained on voc val.
# it achieves 81.0% on voc leaderboard.
# a training script example is as follows.
cd pyramid/VOC/
python resnet-msc-voc-aspp.py   --gpu 0,1,2,3,4,5,6,7  --load [ROOT-TO-LOAD-MODEL]  --data_format NHWC  -d 101  --mode resnet --log_dir [ROOT-TO-SAVE-MODEL]  --data [ROOT-OF-TRAINING-DATA]

• Testing script:

cd pyramid/VOC/
python gr_test_pad_crf_msc_flip.py 

• Trained Models:

  Model trained for evaluation on voc val set:

  Baidu Pan, code: 7dl0

  Google Drive

  Model trained for evaluation on voc leaderboard (test set)

  Baidu Pan, code: 7dl0

  Google Drive

Citation

If you use these models in your research, please cite:

@inproceedings{yang2017learning,
        title={Learning feature pyramids for human pose estimation},
        author={Yang, Wei and Li, Shuang and Ouyang, Wanli and Li, Hongsheng and Wang, Xiaogang},
        booktitle={The IEEE International Conference on Computer Vision (ICCV)},
        volume={2},
        year={2017}
    }

Dependencies

• Python 2.7 or 3
• TensorFlow >= 1.3.0
• Tensorpack The code depends on Yuxin Wu's Tensorpack. For convenience, we provide a stable version 'tensorpack-installed' in this repository.

  # install tensorpack locally:
  cd tensorpack-installed
  python setup.py install --user