Pre-Extracted 2D Features for 3D Instance Segmentation Models

Pre-extracted 2D image-level and object-level features for 3D scenes from the ScanNet and ScanNet200 datasets. The collection contains features for 1513 scenes, with 4539 total files provided by authors Jinyuan Qu, Hongyang Li, and colleagues. The dataset was last updated in February 2026.

Use Cases

• Fine-tune 3D instance segmentation models using pre-computed {scene_id}_query_feats.pth features to reduce training computation.
• Benchmark model performance on ScanNet200 using the provided {scene_id}_query_3dctr.pth object-level features.
• Initialize or augment 3D vision pipelines with the image-level features stored in each {scene_id}.pth file.

Strengths

• Features provided for 1513 distinct 3D scenes.
• Includes 4539 pre-processed feature files, saving computational cost.

Limitations

• Dataset size, row count, and specific feature dimensions are not provided.
• Limited to scenes from the ScanNet and ScanNet200 datasets, which may introduce geographic or scene-type bias.

Provenance

Source

Features extracted from ScanNet and ScanNet200 datasets.

Collection Method

Pre-extracted using the SegDINO3D method to produce image-level and object-level 2D features.

Freshness

Last updated February 2026.