3dlg-hcvc/ReVSI

ICML 2026
Yiming Zhang^1*, Jiacheng Chen^1*, Jiaqi Tan¹, Yongsen Mao², Wenhu Chen³, Angel X. Chang^1,4
1 Simon Fraser University    ² Hong Kong University of Science and Technology
³ University of Waterloo    ⁴ Alberta Machine Intelligence Institute (Amii)

This repository contains the ReVSI benchmark and dataset, introduced in ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning.

Data Subsets

ReVSI provides multiple data subsets corresponding to different video frame budgets:

• all-frame
• 64-frame
• 32-frame
• 16-frame

Use the following command to load a specific subset:

from datasets import load_dataset
revsi_dataset = load_dataset("3dlg-hcvc/ReVSI", "64_frame", split="test")  # load the 64-frame subset

How video subsets are constructed:

The all-frame subset contains the full processed video sequence for each scene, with standardized resolution and frame rate:

1. ScanNet v2 / ScanNetPP v2 / MultiScan
   640 × 480 · 10 FPS

2. ARKitScenes
   640 × 480 / 480 × 640 · 10 FPS (all videos have been rotated to sky-up orientation)

3. 3RScan
   360 × 640 · 4 FPS

The fixed-budget subsets are constructed via hierarchical uniform sampling:

1. Uniformly sample 64 frames from all-frame
2. Uniformly subsample 32 frames from the 64-frame set
3. Uniformly subsample 16 frames from the 32-frame set

This produces a nested structure: 16-frame ⊂ 32-frame ⊂ 64-frame ⊂ all-frame. For each video, all subsets cover the same time span, and each sampled frame keeps the same timestamp across subsets. This guarantees consistent timestamps for models with frame timestamp encoding.

Data Fields

Each entry in ReVSI dataset contains the following fields:

Field Name	Type	Description
id	int64	Unique identifier for each sample
dataset	string	Source dataset of the video
scene_id	string	Identifier of the scene (video) associated with the sample
question_type	string	Category of the question
question	string	Natural language question grounded in the video
options	list[string]	List of answer choices (only for multiple-choice questions)
ground_truth	string	Ground-truth answer to the question
num_frames	string	Frame budget used for evaluation (e.g., 16, 32, 64, all)
queried_object_ids	list[int64]	List of object instance IDs referenced in the question

Evaluation

Please avoid using PyTorch 2.9, as a known cuDNN issue can lead to significant performance degradation for QwenVL models (see details).

ReVSI supports inference / evaluation with the following frameworks:

• LMMs-Eval (inference + evaluation)

  # example 1: evaluate Qwen3-VL-8B-Instruct on ReVSI 64-frame subset (with huggingface transformers backend on 4 GPUs)
  accelerate launch \
  --num_processes=4 \
  -m lmms_eval \
  --model qwen3_vl \
  --model_args=pretrained=Qwen/Qwen3-VL-8B-Instruct,attn_implementation=flash_attention_2,max_num_frames=64 \
  --tasks revsi_64_frame \
  --batch_size 8
  
  # example 2: evaluate Qwen3-VL-8B-Instruct on ReVSI all-frame subset using 2 fps sampling rate (with vllm backend)
  python -m lmms_eval \
  --model vllm \
  --model_args "model=Qwen/Qwen3-VL-8B-Instruct,fps=2" \
  --tasks revsi_all_frame
  

• VLMEvalKit (inference + evaluation)

  # example 1: evaluate Qwen3-VL-8B-Instruct on ReVSI 32-frame subset (with vllm backend)
  python run.py --data revsi_32_frame --model Qwen3-VL-8B-Instruct
  

• ModelScope SWIFT (inference-only, check ReVSI GitHub repo for data registration)

  # example 1: infer Qwen3-VL-8B-Instruct on ReVSI 64-frame subset (with huggingface transformers backend on 4 GPUs)
  NPROC_PER_NODE=4 swift infer \
  --model Qwen/Qwen3-VL-8B-Instruct  \
  --model_kwargs '{"fps_min_frames": 64, "fps_max_frames": 64}' \
  --val_dataset 3dlg-hcvc/ReVSI:64_frame \
  --infer_backend transformers \
  --custom_register_path ./ms_swift_register/revsi_register.py \
  --use_hf true \
  --torch_dtype bfloat16 \
  --attn_impl flash_attention_2 \
  --strict true \
  --max_batch_size 8 \
  --temperature 0
  

• TorchMetrics Extension (evaluation-only)

  # example 1: evaluate existing predictions on ReVSI all-frame subset using TorchMetrics Extension evaluator
  from torchmetrics_ext.metrics.vqa import ReVSIMetric
  
  metric = ReVSIMetric(subset=all_frame)
  predictions = {0: "2", 1: "4", ..., 1000: "A"}  # predictions should be a dict following the format {question_id: response}
  results = metric(pred_dict)
  

Metadata Files

We provide several metadata files used in constructing ReVSI:

• metadata/3d_annotation.json: 3D annotations for each scene, including object names, oriented bounding boxes and scene area polygons. The schema is as follows:

  [
    {
      "scene_id": # scene ID from the source dataset
      "dataset":  # source dataset name
      "scene_area_2d_polygon": # list of 2D boundary points (x, y) defining the scene area polygon, shape (N, 2)
      "scene_area_type":  # scene area annotation type (single_room or multiple_room)
      "objects": [
        {
          "id":  # object id within the scene
          "name":  # open-vocabulary object name
          "obb": {
            "center":  # center of the object oriented bounding boxes, shape (3, )
            "extent":  # extent of the object oriented bounding boxes, shape (3, )
            "rotation":  # rotation matrix of the object oriented bounding boxes, shape (3, 3)
          }
        },
        ...
      ]
    },
    ...
  ]
  

• metadata/sampled_video_frame_idx.json: indices of sampled frames for the 16/32/64-frame subsets. The scehema is as follows:

  {
    "<scene_id>": {
      "64-frame":  # list of sampled frame indices from the all-frame video, shape (64, )
      "32-frame":  # list of sampled frame indices from the all-frame video, shape (32, )
      "16-frame":  # list of sampled frame indices from the all-frame video, shape (16, )
    }
    ...
  }
  

• metadata/obj_visibility.json: Object visibility under different video frame budgets. The schema is as follows:

  {
    "<scene_id>": [
      {
        "object_id":  # object id within the scene (consistent with metadata/3d_annotation.json)
        "object_name":  # open-vocabulary object name (consistent with metadata/3d_annotation.json)
        "visibility_16":  # visibility under the 16-frame budget
        "visibility_32":  # visibility under the 32-frame budget
        "visibility_64":  # visibility under the 64-frame budget
      },
      ...
    ],
    ...
  }
  

• metadata/tiny_set_question_ids.txt: The sampled question ids of tiny set for proprietary model evaluations.

Citation

If you find ReVSI useful for your research, please consider citing:

@article{zhang2026revsi,
  title={ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning},
  author={Zhang, Yiming and Chen, Jiacheng and Tan, Jiaqi and Mao, Yongsen and Chen, Wenhu and Chang, Angel X.},
  journal={arXiv preprint arXiv:2604.24300},
  year={2026}
}

ReVSI builds upon the following 3D scene datasets and the VSI-Bench benchmark, please also consider citing:

@inproceedings{dai2017scannet,
  title={Scannet: Richly-annotated 3d reconstructions of indoor scenes},
  author={Dai, Angela and Chang, Angel X and Savva, Manolis and Halber, Maciej and Funkhouser, Thomas and Nie{\ss}ner, Matthias},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={5828--5839},
  year={2017}
}

@inproceedings{yeshwanth2023scannet++,
  title={Scannet++: A high-fidelity dataset of 3d indoor scenes},
  author={Yeshwanth, Chandan and Liu, Yueh-Cheng and Nie{\ss}ner, Matthias and Dai, Angela},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={12--22},
  year={2023}
}

@inproceedings{baruch1arkitscenes,
  title={ARKitScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data},
  author={Baruch, Gilad and Chen, Zhuoyuan and Dehghan, Afshin and Feigin, Yuri and Fu, Peter and Gebauer, Thomas and Kurz, Daniel and Dimry, Tal and Joffe, Brandon and Schwartz, Arik and others},
  booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1)}
}

@inproceedings{wald2019rio,
  title={Rio: 3d object instance re-localization in changing indoor environments},
  author={Wald, Johanna and Avetisyan, Armen and Navab, Nassir and Tombari, Federico and Nie{\ss}ner, Matthias},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={7658--7667},
  year={2019}
}

@article{mao2022multiscan,
  title={Multiscan: Scalable rgbd scanning for 3d environments with articulated objects},
  author={Mao, Yongsen and Zhang, Yiming and Jiang, Hanxiao and Chang, Angel and Savva, Manolis},
  journal={Advances in neural information processing systems},
  volume={35},
  pages={9058--9071},
  year={2022}
}

@inproceedings{yang2025thinking,
  title={Thinking in space: How multimodal large language models see, remember, and recall spaces},
  author={Yang, Jihan and Yang, Shusheng and Gupta, Anjali W and Han, Rilyn and Fei-Fei, Li and Xie, Saining},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={10632--10643},
  year={2025}
}