3DModelSearch — Renders & Embeddings

Precomputed data for 3DModelSearch: multi-view image renders, turntable videos, and LanguageBind embeddings for 3D models, used in Learning Visual and Motion Aware 3D Model Representations for Semantic Retriev. These let you reproduce retrieval results without re-rendering or re-embedding.

Contents

File	Size	Contents
embeddings.tar.gz	~ GB	out/Embeddings/<uid>_image.pt, <uid>_video.pt, and out/models_data.csv
videos.tar.gz	~17 GB	data/Videos/<shard>/<uid>.mp4
images.tar.gz	~197 GB	data/Images/<shard>/<uid>/view0..5.jpg

• Image embeddings: [6, 768] tensors (6 views, LanguageBind image encoder).
• Video embeddings: [768] tensors (LanguageBind video encoder).
• models_data.csv maps each uid to its render + embedding paths (paths are relative to the code repo root).

Source .glb models are not included — fetch them from Objaverse as described in the code repo.

Usage

pip install huggingface_hub
huggingface-cli download <HF_USER>/<REPO> embeddings.tar.gz --repo-type dataset --local-dir .
tar -xzf embeddings.tar.gz     # -> out/
tar -xzf videos.tar.gz         # -> data/Videos/
tar -xzf images.tar.gz         # -> data/Images/