ai-jie01/vl-plm

Exploiting unlabeled data with vision and language models for object detection, ECCV 2022

Exploiting Unlabeled Data with Vision and Language Models for Object Detection

Official implementation of Exploiting unlabeled data with vision and language models for object detection.

arXiv, Project

Installation

Our project is developed on Detectron2. Please follow the official installation instructions.

Data Preparation

Download the COCO dataset, and put it in the datasets/ directory.

Download our pre-generated pseudo-labeled data, and put them in the datasets/open_voc directory.

Dataset are organized in the following way:

datasets/
    coco/
        annotations/
            instances_train2017.json
            instances_val2017.json
            open_voc/
                instances_eval.json
                instances_train.json
        images/
            train2017/
                000000000009.jpg
                000000000025.jpg
                ...
            val2017/
                000000000776.jpg
                000000000139.jpg
                ...
        

Note: You may generate and evaluate pseudo labels on your own by following our pseudo label generation instruction

Evaluation with pre-trained models

Mask R-CNN:

Training Method	Novel AP	Base AP	Overall AP	download
With LSJ	34.4	60.2	53.5	model
W/O LSJ	32.3	54.0	48.3	model

python -m train_net.py --config configs/coco_openvoc_LSJ.yaml  --num-gpus=1 --eval-only --resume

Training

The best model on COCO in the paper is trained with large scale Jitter (LSJ), but training with LSJ requires too many GPU memories. Thus, beside the LSJ version, we also provide training without LSJ.

Training Mask R-CNN with Large Scale Jitter (LSJ).

python train_net.py --config configs/coco_openvoc_LSJ.yaml  --num-gpus=8 --use_lsj

Training Mask R-CNN without Large Scale Jitter (LSJ).

python train_net.py --config configs/coco_openvoc_mask_rcnn.yaml  --num-gpus=8

Citing VL-PLM

If you use VL-PLM in your work or wish to refer to the results published in this repo, please cite our paper: