PSOL

We have updated the training and validation code for PSOL on ImageNet;

This is the offical website for PSOL.

We have uploaded the poster for PSOL. This package is developed by Mr. Chen-Lin Zhang (http://www.lamda.nju.edu.cn/zhangcl/) and Mr. Yun-Hao Cao (http://www.lamda.nju.edu.cn/caoyh/). If you have any problem about the code, please feel free to contact Mr. Chen-Lin Zhang ([email protected]). The package is free for academic usage. You can run it at your own risk. For other purposes, please contact Prof. Jianxin Wu (mailto:[email protected]).

If you find our package is useful to your research, please cite our paper:

Reference:

[1] C.-L. Zhang, Y.-H. Cao and J. Wu. Rethinking the Route Towards Weakly Supervised Object Localization . In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2020), Seattle, WA (Virtual), USA, pp. 13460-13469.

Requirements

The code needs PyTorch and OpenCV.

A GPU is optional for faster speed.

Other requirements are in requirements.txt. You can simply install them by pip install -r requirements.txt. We recommend to use Anaconda with virtual environments.

Usage

prepare datasets:

Training:

First you should download the ImageNet training set. Then, use our program generate_box_imagenet.py to generate pseudo bounding boxes by DDT methods:

python generate_box_imagenet.py PATH

We follow the organization of PyTorch official ImageNet training script. There are some options in generate_box_imagenet.py:

PATH            the path to the ImageNet dataset
--input_size    input size for each image (448 for default)
--gpu           which gpus to use 
--output_path   the output pseudo boxes folder
--batch_size    batch_size for executing forward pass of CNN

If you use default parameters, you are just using DDT-VGG16 with 448x448 as illustrated in the paper. For efficiency, we provide generated VGG-16 448x448 boxes pseudo boxes. You can directly download it.

Validation:

First you should download the ImageNet validation set, and corresponding annotation xmls. We need both validation set and annotation xmls in PyTorch format. Please refer to ImageNet example for details.

Folder Structure

We except the ImageNet Folder has these structures:

root/
|-- train/
|   |-- class1 |-- image1.jpg 
|   |-- class2 |-- image2.jpg
|   |-- class3 |-- image3.jpg
|   ...
|-- val/
|   |-- class1 |-- image1.jpg 
|   |-- class2 |-- image2.jpg
|   |-- class3 |-- image3.jpg
|   ...
|-- myval/ (groundtruth annotation xml file, you can change the folder name, and modify it in Line 67 in PSOL_inference.py)
|   |-- class1 |-- image1.xml 
|   |-- class2 |-- image2.xml
|   |-- class3 |-- image3.xml
...

Training

Please first refer to prepare datasets session for generating pseudo bounding boxes. Then, you can use PSOL_training.py to train PSOL-Sep models on pseudo bounding boxes.

python PSOL_training.py PATH. There are some options in PSOL_training.py:

PATH            the path to the ImageNet dataset
--loc-model     which localization model to use. Current options: densenet161, vgg16,vgggap, resnet50, densenet161
--input_size    input size for each image (256 for default)
--crop_size     crop size for each image (only used in validation process)
--ddt_path      the path of generated DDT pseudo boxes
--gt_path       the path for groundtruth on ImageNet val dataset
--save_path     where to save the checkpoint
--gpu           which gpus to use 
--batch_size    batch_size for executing forward pass of CNN

Testing

We have provided some pretrained models and cached groundtruth files. If you want to directly test it, please download it.

We use same options as PSOL_training.py in PSOL_inference.py. Then you can run:

python PSOL_inference.py --loc-model {$LOC_MODEL} --cls-model {$CLS_MODEL} {--ten-crop}

Extra Options:

--cls-model     represents the classification models we support now. Current options: densenet161, vgg16, resnet50, inceptionv3, dpn131, efficientnetb7 (need to install efficientnet-pytorch). 
--ten-crop      use ten crop to boost the classification performance.

Please note that for SOTA classification models, you should change the resolutions in cls_transforms (Line248-Line249).

Then you can get the final Corloc and Clsloc result.