SNARE Dataset

March 27, 2024 · View on GitHub

SNARE dataset and code for MATCH and LaGOR models.

Paper and Citation

@article{snare,
  title={Language Grounding with {3D} Objects},
  author={Jesse Thomason and Mohit Shridhar and Yonatan Bisk and Chris Paxton and Luke Zettlemoyer},
  journal={arXiv},
  year={2021},
  url={https://arxiv.org/abs/2107.12514}
}

Installation

Clone

$ git clone https://github.com/snaredataset/snare.git

$ virtualenv -p $(which python3) --system-site-packages snare_env # or whichever package manager you prefer
$ source snare_env/bin/activate

$ pip install --upgrade pip
$ pip install -r requirements.txt

Edit root_dir in cfgs/train.yaml to reflect your working directory.

Download Data and Checkpoints

Download pre-extracted image features, language features, and pre-trained checkpoints from here and put them in the data/ folder.

Usage

Zero-shot CLIP Classifier

$ python train.py train.model=zero_shot_cls train.aggregator.type=maxpool

MATCH

$ python train.py train.model=single_cls train.aggregator.type=maxpool

LaGOR

$ python train.py train.model=rotator train.aggregator.type=two_random_index train.lr=5e-5 train.rotator.pretrained_cls=<path_to_pretrained_single_cls_ckpt>

Scripts

Run scripts/train_classifiers.sh and scripts/train_rotators.sh to reproduce the results from the paper.

To train the rotators, edit scripts/train_rotators.sh and replace the PRETRAINED_CLS with the path to the checkpoint you wish to use to train the rotator:

PRETRAINED_CLS="<root_path>/clip-single_cls-random_index/checkpoints/<ckpt_name>.ckpt'"

Preprocessing

If you want to extract CLIP vision and language features from raw images:

Download models-screenshot.zip from ShapeNetSem, and extract it inside ./data/.
Edit and run python scripts/extract_clip_features.py to save shapenet-clipViT32-frames.json.gz and langfeat-512-clipViT32.json.gz

Leaderboard

Please send your ...test.json prediction results to Mohit Shridhar. We will get back to you as soon as possible.

Instructions:

Include a name for your model, your team name, and affiliation (if not anonymous).
Submissions are limited to a maximum of one per week. Please do not create fake email accounts and send multiple submissions.

Rankings:

Rank	Model	All	Visual	Blind
1	DA4LG (Anonymous) 5 Feb 2024	81.9	88.5	75.0
2	MAGiC (Mitra et al.) 8 Jun 2023	81.7	87.7	75.4
3	DA4LG (Anonymous) 27 Jan 2024	80.9	87.7	73.7
4	VLG (Corona et al.) 15 Mar 2022	79.0	86.0	71.7
5	LOCKET (Anonymous) 14 Oct 2022	79.0	86.1	71.5
6	VLG (Corona et al.) 13 Nov 2021	78.7	85.8	71.3
7	LOCKET (Anonymous) 23 Oct 2022	77.7	85.5	69.5
8	LAGOR (Thomason et. al) 15 Sep 2021	77.0	84.3	69.4
9	MATCH (Thomason et. al) 15 Sep 2021	76.4	83.7	68.7