3D-Point-QA

3D-Point-QA is a low-level spatial reasoning benchmark introduced in Render to Reason: Novel-View Semantic Prediction Improves Spatial Understanding in VLMs.

[Project Page] [Paper] [Code]

Existing spatial benchmarks refer to objects only by name, so a wrong answer can come from either geometric reasoning or object grounding. 3D-Point-QA instead marks points with colored arrows on one to four views, and asks questions that are straightforward to answer given camera poses and point clouds, without object-level reasoning. Answers are computed from the ground-truth meshes and camera poses of ScanNet++, following its official splits.

Tasks

Task (task_type)	Question	Metric
Point-to-camera distance (distance_to_camera)	Distance from each marked point to the camera	δ1.25
Point-to-point distance (distance_prediction)	3D distance between two marked points	δ1.25
Relative distance (distance_infer)	Which of two points is closer to a reference point	accuracy
Point matching (position_matching)	Where a marked point appears in another view	PCK@0.05
3D coordinate mapping (spatial_imagination_3d)	3D coordinates of the marked points in a common frame	% within 0.5 m
Object counting (counting, train only)	Number of distinct objects of a class per view and in total	—

Splits

Split	Scenes	Questions	Images
train	855 (ScanNet++ nvs_sem_train)	97,999 (17,955 per point task + 8,224 counting)	90,767
validation	50 (ScanNet++ nvs_sem_val)	5,250 (1,050 per point task)	5,098

Scenes that appear in VSI-Bench are excluded from training.

Files

train.jsonl          train_images.tar   -> images/<scene>/frame_XXXXXX.jpg
val.jsonl            val_images.tar     -> images/<scene>/frame_XXXXXX.jpg

hf download yuqun/3D-Point-QA --repo-type dataset --local-dir 3d_point_qa
cd 3d_point_qa && tar -xf val_images.tar   # and train_images.tar for training

Image paths in the jsonl files are relative to the dataset root.

Format

Each line is one question:

{
  "id": "09c1414f1b_distance_to_camera_23706972",
  "scene_id": "09c1414f1b",
  "task_type": "distance_to_camera",
  "images": ["images/09c1414f1b/frame_004980.jpg"],
  "points": [
    {"color": "red", "label": "surfboard", "obj_id": 188, "img_idx": 0, "norm_xy": [75, 262]},
    ...
  ],
  "conversations": [
    {"from": "human", "value": "In the image, colored arrows indicate points on three objects: ..."},
    {"from": "gpt", "value": "Red: 3.4\nGreen: 3.1\nBlue: 5.3"}
  ],
  "ground_truth_dists_m": [3.41, 3.06, 5.28]
}

• points: the marked points. img_idx is the index into images, and norm_xy is the pixel position scaled to [0, 1000] (x from the left, y from the top).
• The images are stored without arrows. The arrows are drawn at load time from points (see draw_arrows_on_image in the code), so you can render them at any size.
• Task-specific ground truth: ground_truth_dists_m (point-to-camera), distance_m (point-to-point), closer_color, d_red_green and d_red_blue (relative distance), answer_norm_xy and answer_img_idx (point matching), positions_m and world_points (3D mapping), count, per_frame_counts, label and n_views (counting).

Evaluation

cd R2R/qwen-vl-finetune
CKPT=checkpoints/R2R-Qwen3-VL-8B GPUS="0 1 2 3" bash scripts/eval_3d_point_qa.sh

License

The images and annotations are derived from ScanNet++ and are subject to the ScanNet++ Terms of Use.

Citation

@article{wu2026render,
  title   = {Render to Reason: Novel-View Semantic Prediction Improves Spatial Understanding in VLMs},
  author  = {Wu, Yuqun and Xiao, Yao and Zou, Chuhang and Wang, Shenlong and Hoiem, Derek},
  journal = {arXiv preprint arXiv:2610.05417},
  year    = {2026}
}