Yiming1234/psedud

BAGEL-SAM3D training datasets

This repository contains the crafted 3D VQA corpus, camera-up supervision, natural VLM replay, and the separate camera-up step-2,000 checkpoint.

Crafted 3D VQA: pose-grounding v11

data/pose_grounding_v11 contains 119,011 training records:

Source	Records
Indoor (ARKitScenes / HyperSim)	56,796
nuScenes	56,521
KITTI	275
COCO3D	5,419

The four dataset_*.jsonl files are byte-for-byte copies of the training sources. They include crafted spatial questions, ordered answer choices, answers, natural object descriptions, SAM3D tool requests and observations, teacher reasoning, camera metadata, and pose targets. The 15 relation types include relative height, distance, location, and orientation. Images, instance masks, and cached aligned pointmaps used by the training loader are included in six asset archives.

The v11.1 recipe uses this same v11 corpus. Its loader changes the sampling and answer supervision and rotates answer choices for selected relations. Both exact training configs and original launchers are in provenance/pose_grounding_v11. See dataset_summary.json for counts and release_manifest.json for checksums.

Download and prepare

hf download Yiming1234/bagel-sam-training --repo-type dataset \
  --local-dir /path/to/bagel-sam-training
python /path/to/bagel-sam-training/scripts/prepare_released_3d_vqa.py \
  /path/to/bagel-sam-training --extract
tar -xf /path/to/bagel-sam-training/data/llava_onevision_replay/images.tar \
  -C /path/to/bagel-sam-training/data/llava_onevision_replay

The helper extracts 16,240 images, 40,167 masks, and 10,481 pointmaps, validates asset sizes, and writes local training JSONLs and absolute manifests under .local/pose_grounding_v11. Published source files are preserved. Some unused provenance fields retain the original paths to source annotations and teacher models; those are historical metadata, not training asset dependencies.

Set BAGEL_SAM3D_SINGLE_TRIGGER_TEACHER_MANIFEST to the generated manifest_dataset_info.json, and BAGEL_LLAVA_ONEVISION_REPLAY_MANIFEST to the generated replay_manifest.json. Code, dependencies, and a portable v11.1 launcher are at plzdoo/bagel-sam. Reproducing the v11.1 run also requires its parent v11 step-25,000 checkpoint, which is not included in this dataset release.

Camera-up and replay

• data/camera_up_v1: 6,938 training records; 708 validation and 718 test records, plus rotated versions of validation/test. Extract its images.tar into that directory.
• data/llava_onevision_replay: 137,224 natural VLM replay records after excluding 745 records overlapping 3DSRBench COCO image IDs.
• checkpoints/camera_up_v1_step_2000: the separate camera-up checkpoint. It is not the crafted 3D VQA model checkpoint.

The camera-up run samples camera-up/replay at 85%/15%. The v11 VQA mixture samples indoor/outdoor/COCO3D/replay at 40%/30%/15%/15%; v11.1 samples the three VQA supervision modes/replay at 15%/30%/45%/10%.

Images derive from their respective source datasets, including Omni3D/OmniNOCS sources, COCO3D/COCO, and LLaVA OneVision. This release does not grant additional rights to those source assets; retain their original terms and attribution.