[Benchmark] Add Spatial-DISE benchmark

[shinmohuang]

Summary

This PR adds Spatial-DISE, an ICLR 2026 spatial reasoning benchmark, to VLMEvalKit.

Spatial-DISE evaluates VLMs on 2D and 3D spatial reasoning tasks, including rotation, folding, projection, shape finding, and compositional reasoning. The default dataset entry follows the official benchmark split:

DISE-bench/DISE-benchmark.csv

Supported entries:

• Spatial-DISE
• Spatial-DISE_BENCH

Additional reusable dataset splits are also exposed:

• Spatial-DISE_TRAIN
• Spatial-DISE_VAL
• Spatial-DISE_TEST

Implementation

This PR follows the VLMEvalKit benchmark contribution pattern by adding a dataset class with:

• build_prompt(self, line)
• evaluate(self, eval_file, **judge_kwargs)

The implementation:

• adds vlmeval/dataset/spatial_dise.py
• registers SpatialDISE in vlmeval/dataset/__init__.py
• loads the official dataset from Hugging Face:
  TACPS-liv/Spatial-DISE
• supports local reuse via:
  SPATIAL_DISE_ROOT=/path/to/Spatial-DISE
• reads images from tar shards on demand instead of requiring users to manually extract all images
• maps CSV paths like images/... to tar members by removing the images/ prefix
• reports accuracy breakdowns by:
  • category
  • difficulty
  • dise_category

Why This Format

Spatial-DISE is distributed on Hugging Face with CSV metadata and tar-sharded images. Instead of embedding images into a large TSV/base64 file, this PR keeps the official dataset layout and provides a VLMEvalKit-compatible loader that handles tar-shard image access transparently.

Users can run the benchmark directly through the standard VLMEvalKit interface:

python run.py --data Spatial-DISE --model <MODEL_NAME>

Validation

Validated locally with:

python3.10 -m py_compile vlmeval/dataset/spatial_dise.py vlmeval/dataset/__init__.py
git diff --check

Smoke test with:

SPATIAL_DISE_ROOT=/LOCAL2/hxm826/Spatial-DISE-hf

confirmed that:

• Spatial-DISE loads 559 official benchmark examples
• both 2D and 3D categories are present
• benchmark images are extracted from tar shards on demand
• exact-matching evaluation on 12 gold predictions returns 100% overall accuracy