JAEGER

This repository contains the JAEGER model checkpoints and the SpatialSceneQA dataset archives.

TODO before final public release: replace this placeholder card with the paper title, citation, license terms, intended use, and dataset/model access policy.

Repository id: tsinghua-ee/JAEGER

Repository type used by the upload script: model. The default is a model repository so the JAEGER checkpoints are the primary artifact; the SpatialSceneQA archives are stored under datasets/SpatialSceneQA/.

Directory Structure

datasets/SpatialSceneQA/
  hm3d_foa_av_v2_train.tar.gz
  hm3d_foa_av_v2_val.tar.gz
  hm3d_foa_av_v2_test.tar.gz
checkpoints/
  classical_iv/
    audio_only_1/
    audio_only_2/
    audio_visual_1/
    audio_visual_2/
    visual_grounding/v_o/
    visual_grounding/v_o_spk/
  neural_iv/
    audio_only_1/
    audio_only_2/
    audio_visual_1/
    audio_visual_2/

neural_iv/visual_grounding/README.md notes that the visual grounding checkpoint is shared with classical_iv, following the source checkpoint organization.

Download Examples

Install dependencies:

pip install -U huggingface_hub hf_xet

Download only the validation split:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="tsinghua-ee/JAEGER",
    repo_type="model",
    filename="datasets/SpatialSceneQA/hm3d_foa_av_v2_val.tar.gz",
)
print(path)

Download all classical IV checkpoints:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="tsinghua-ee/JAEGER",
    repo_type="model",
    allow_patterns="checkpoints/classical_iv/**",
    local_dir="JAEGER-classical-iv",
)

Download everything:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="tsinghua-ee/JAEGER",
    repo_type="model",
    local_dir="JAEGER",
)