VLM-3R Training Data

Training QA data for VLM-3R: vsibench_train/ (VSI-Bench-style tasks) and vstibench_train/ (VSTI-Bench tasks over ScanNet train split).

Erratum (2026-07-13): corrected camera-position ground truth

A bug in the QA generation pipeline (reported by Jacob Yeung, CMU) extracted the camera center from camera-to-world poses using -R.T @ t instead of pose[:3, 3]. Answers in five vstibench_train files depended on the camera's world position and have been corrected in place; items whose corrected geometry violates the generators' validity filters (distance range, option ambiguity margin, minimum object distance) were removed and are listed in vstibench_train/erratum_dropped_items.json.

Affected files: qa_camera_displacement.json, qa_camera_obj_abs_dist.json, qa_camera_obj_rel_dist_v1/v2/v3.json. All other files (camera movement direction, object-object relative position, and all vsibench_train files) do not depend on the camera position bug and are unchanged.

Models fine-tuned on the pre-correction files learned a systematically biased target for these question types; retraining on this revision is recommended. Pre-correction files remain available in the git history. The corrected generation code and correction scripts are in the VLM-3R repository (vlm_3r_data_process/scripts/vstibench/).

The corresponding evaluation benchmark was corrected in the same way; see the erratum in Journey9ni/vstibench.