MindView-Bench

MindView benchmarks language-guided target-view generation from multiple observations of an indoor scene. Models combine reference RGB images with a spatial instruction to generate the requested viewpoint while preserving scene geometry and appearance.

The official test set contains 1,500 examples from 513 scenes across ScanNet, ScanNet++, and ARKitScenes. Separate context-scaling experiments reuse these same examples.

Code · Training data · BAGEL-SFT weights · Collection

Contents

Path	Purpose
data/test.jsonl	Official 1,500-example test index: ordered generator inputs and instructions
context_scaling/	Additional C3/C6/C10 experiments: 4,500 aligned configuration records
evaluation/references.jsonl	Fixed scoring references, held-out targets, depth, and source paths
evaluation/protocol.json	Tasks, metric definitions, and aggregation rules
assets/{rgb,depth}/*.tar	Original image bytes in 20 ordinary TAR archives
manifests/	Per-asset and per-archive paths, sizes, and checksums
scripts/prepare_data.py	Verify the download and safely extract assets
RELEASE.json, SHA256SUMS	Release inventory and file checksums

The archives contain 14,276 RGB images and 14,276 depth maps, deduplicated across test inputs, scoring references, targets, and additional experiments. Archive counts do not indicate images per example.

Generator inputs and scoring

Official generation uses six ordered context RGB images and the instruction in each test record. VGGT scoring uses a fixed set of ten reference RGB images, plus the held-out target and generated image: twelve RGB images in total. Sensor depth supports metric calibration. The extra scoring references, target RGB, and depth must not enter the generator beyond its prescribed inputs.

Seven tasks cover translation, rotation, composition, relational object grounding, target-facing reorientation, object-centric repositioning, and multi-cue viewpoint inference. Six primary metrics—DINO-I, Semantic, Spatial, Quality, VGGT-RScore, and VGGT-TScore—are averaged within each task, then equally across tasks. Overall averages the six resulting scores; auxiliary CLIP-I is excluded. Report context-scaling results separately.

The frozen subset was selected using model-level diagnostics; results for methods involved in that selection are selection-conditioned.

Download and prepare

The first complete release is tagged v1.0. Use the same revision for annotations and assets. Install the Hugging Face CLI if needed:

pip install -U huggingface_hub
hf download wangyanbo/MindView-Bench \
  --repo-type dataset --revision "v1.0" \
  --local-dir ./MindView-Bench

With Python 3.10 or later, verify the complete snapshot, then extract into a separate directory:

python MindView-Bench/scripts/prepare_data.py \
  --repo MindView-Bench --verify-only
python MindView-Bench/scripts/prepare_data.py \
  --repo MindView-Bench --output ./mindview-data

Extraction restores mindview-data/scannet/, scannetpp/, and arkitscenes/. Matching existing files are reused; conflicting files and unsafe archive entries are rejected. Images and depth values are preserved. Depth is stored as unsigned 16-bit PNG in millimeters; evaluation converts valid values to meters.

For annotations only:

from datasets import load_dataset

test = load_dataset(
    "wangyanbo/MindView-Bench", "benchmark",
    split="test", revision="v1.0",
)
assert len(test) == 1500

Annotations contain frame IDs and paths; TAR assets are downloaded and extracted separately rather than streamed as decoded Image features.

Runtime configuration

In the project runtime, explicitly set INPUT_JSON to MindView-Bench/data/test.jsonl and the three dataset roots to the extracted source directories. For VGGT, set VGGT_CANONICAL_JSON to MindView-Bench/context_scaling/metadata.jsonl; the scorer selects its ten-view records. evaluation/references.jsonl documents the scoring assets but is not automatically consumed by the current scorer.

Older launchers default to the combined 4,500-record index. Do not use that default for official generation. Package integrity has been checked; GPU evaluation with this release layout still requires validation.

Terms and contact

Source-data terms remain applicable; this package does not relicense ScanNet, ScanNet++, or ARKitScenes. Read DATA_LICENSE.md before use.

Contact: Yihao Liu (liuyihao@pjlab.org.cn) and Jiwen Lu (lujiwen@tsinghua.edu.cn). Include the revision and sample ID when reporting issues.