jiachen0212/3D-VisTA

Official implementation of ICCV 2023 paper "3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment"

3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment

Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng📧, Siyuan Huang📧, Qing Li📧

This repository is the official implementation of the ICCV 2023 paper "3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment".

Paper | arXiv | Project | HuggingFace Demo | Checkpoints

Abstract

3D vision-language grounding (3D-VL) is an emerging field that aims to connect the 3D physical world with natural language, which is crucial for achieving embodied intelligence. Current 3D-VL models rely heavily on sophisticated modules, auxiliary losses, and optimization tricks, which calls for a simple and unified model. In this paper, we propose 3D-VisTA, a pre-trained Transformer for 3D Vision and Text Alignment that can be easily adapted to various downstream tasks. 3D-VisTA simply utilizes self-attention layers for both single-modal modeling and multi-modal fusion without any sophisticated task-specific design. To further enhance its performance on 3D-VL tasks, we construct ScanScribe, the first large-scale 3D scene-text pairs dataset for 3D-VL pre-training. ScanScribe contains 2,995 RGB-D scans for 1,185 unique indoor scenes originating from ScanNet and 3R-Scan datasets, along with paired 278K scene descriptions generated from existing 3D-VL tasks, templates, and GPT-3. 3D-VisTA is pre-trained on ScanScribe via masked language/object modeling and scene-text matching. It achieves state-of-the-art results on various 3D-VL tasks, ranging from visual grounding and dense captioning to question answering and situated reasoning. Moreover, 3D-VisTA demonstrates superior data efficiency, obtaining strong performance even with limited annotations during downstream task fine-tuning.

Install

1. Install conda package

conda env create --name 3dvista --file=environments.yml

2. install pointnet2

cd vision/pointnet2
python3 setup.py install

Prepare dataset

1. Follow Vil3dref and download scannet data under data/scanfamily/scan_data, this folder should look like

./data/scanfamily/scan_data/
├── instance_id_to_gmm_color
├── instance_id_to_loc
├── instance_id_to_name
└── pcd_with_global_alignment

2. Download scanrefer+referit3d, scanqa, and sqa3d, and put them under /data/scanfamily/annotations

data/scanfamily/annotations/
├── meta_data
│   ├── cat2glove42b.json
│   ├── scannetv2-labels.combined.tsv
│   ├── scannetv2_raw_categories.json
│   ├── scanrefer_corpus.pth
│   └── scanrefer_vocab.pth
├── qa
│   ├── ScanQA_v1.0_test_w_obj.json
│   ├── ScanQA_v1.0_test_wo_obj.json
│   ├── ScanQA_v1.0_train.json
│   └── ScanQA_v1.0_val.json
├── refer
│   ├── nr3d.jsonl
│   ├── scanrefer.jsonl
│   ├── sr3d+.jsonl
│   └── sr3d.jsonl
├── splits
│   ├── scannetv2_test.txt
│   ├── scannetv2_train.txt
│   └── scannetv2_val.txt
└── sqa_task
    ├── answer_dict.json
    └── balanced
        ├── v1_balanced_questions_test_scannetv2.json
        ├── v1_balanced_questions_train_scannetv2.json
        ├── v1_balanced_questions_val_scannetv2.json
        ├── v1_balanced_sqa_annotations_test_scannetv2.json
        ├── v1_balanced_sqa_annotations_train_scannetv2.json
        └── v1_balanced_sqa_annotations_val_scannetv2.json

3. Download all checkpoints and put them under project/pretrain_weights

Checkpoint	Link	Note
Pre-trained	link	3D-VisTA Pre-trained checkpoint.
ScanRefer	link	Fine-tuned ScanRefer from pre-trained checkpoint.
ScanQA	link	Fine-tined ScanQA from pre-trained checkpoint.
Sr3D	link	Fine-tuned Sr3D from pre-trained checkpoint.
Nr3D	link	Fine-tuned Nr3D from pre-trained checkpoint.
SQA	link	Fine-tuned SQA from pre-trained checkpoint.
Scan2Cap	link	Fine-tuned Scan2Cap from pre-trained checkpoint.

Run 3D-VisTA

To run 3D-VisTA, use the following command, task includes scanrefer, scanqa, sr3d, nr3d, sqa, and scan2cap.

python3 run.py --config project/vista/{task}_config.yml

Acknowledgement

We would like to thank the authors of Vil3dref and for their open-source release.

News

• [ 2023.08 ] First version!
• [ 2023.09 ] We release codes for all downstream tasks.

Citation:

@article{zhu2023vista,
  title={3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment},
  author={Zhu, Ziyu and Ma, Xiaojian and Chen, Yixin and Deng, Zhidong and Huang, Siyuan and Li, Qing},
  journal={ICCV},
  year={2023}
}