Awesome

MarvelOVD: Marrying Object Recognition and Vision-Language Models for Robust Open-Vocabulary Object Detection

Official implementation of MarvelOVD in ECCV 2024.

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 VL-PLM pre-generated pseudo-labeled data and our generated candidate pseudo-label 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
                ...

Pseudo label generation

MarvelOVD dynamically learns open-vocabulary knowledge from offline-generated pseudo-labels under the guidance from the online training detector.

If necessary, please refer to pseudo label generation instruction to generate offline pseudo-labels.

Results on OVD-COCO

Mask R-CNN:

<table><tbody>   <th valign="bottom">Novel AP</th> <th valign="bottom">Base AP</th> <th valign="bottom">Overall AP</th>   <tr> <td align="center">38.9</td> <td align="center">56.4</td> <td align="center">51.8</td> </tr> </tbody></table>

Training

We train the model under regular data augmentations (no Large Scale Jittering), without extra GPU memory occupation. (Runing on 4 GPUs with 24G Memory per GPU)

Training command

python train_net.py --config configs/coco_ssod.yaml  --num-gpus=4

Notes

The code is highly borrowed from VL_PLM, big thanks for the open-source commuity. Questions and Issues, please contract wangk229@mail2.sysu.edu.cn