DuplicateSingleImage

Dataset for Structure from Duplicates: Neural Inverse Graphics from a Pile of Objects (NeurIPS 2023). Each object is a single image of N identical instances, treated as N virtual views of one object, plus the poses/point cloud recovered from it by the SfD preprocessing pipeline.

• Code: https://github.com/tianhang-cheng/SfD
• Blender source scenes: TianhangCheng7/DuplicateBlenderData
• Pretrained weights / SuperGlue checkpoints: TianhangCheng7/DuplicateWeight

Download

pip install -U huggingface_hub

# everything
hf download TianhangCheng7/DuplicateSingleImage --repo-type dataset --local-dir DuplicateSingleImage

# one object is enough to try training
hf download TianhangCheng7/DuplicateSingleImage --repo-type dataset \
    --include "train_split/coffee/*" "eval_split/coffee/*" --local-dir DuplicateSingleImage

Then train directly out of the download — no copying required:

python exp_runner.py --conf configs/default.yaml \
    --data_split_dir DuplicateSingleImage/train_split/coffee \
    --expname coffee --trainstage Geo --init_method SFM

Layout

train_split/<object>/
  train/                        # 800x800, what training reads
    000_rgb.png | 000_rgb.exr   # .exr for the synthetic objects
    000_instance_seg.png        # 0 = background, i/N*255 = instance i
    000_normal_pretrain.png     # Omnidata monocular normal prior
    000_normal.png              # synthetic only (renderer GT)
    000_diffuse.png             # synthetic only
    000_roughness.png           # synthetic only
  highres_for_matching/         # the preprocessing INPUT (see below)
    000_rgb.png                 # 3072x3072 or 3200x3200, 8-bit
    000_instance_seg.png
  transforms_train.json         # virtual camera intrinsics/extrinsics
  object_pred_pose.json         # per-instance pose recovered by SfM
  object_scale_matrix.json
  points_world.npy              # sparse SfM point cloud, per instance
  non_empty_indexes.txt         # instances COLMAP managed to register
  blender_object_gt_pose.json   # synthetic only, GT pose for evaluation
  blender_camera_gt_pose.json   # synthetic only

eval_split/<object>/
  transforms_test.json
  depth_sfm_bar_origin.png
  train/000_mask.png
  train/000_diffuse.png                 # synthetic only, albedo GT
  train/000_roughness.png               # synthetic only
  train/000_metallic.png                # synthetic only
  test_relight_b/ test_relight_d/       # only objects with relighting GT

15 objects: airplane, box, cake, cash, cheese, cleaner, clock, coffee, cola, fire, gitar, potato, sign, tin, yogurt.

highres_for_matching

train/ holds the 800×800 images training consumes — these are the output resolution of preprocessing and are too small to re-derive their own annotations (SuperPoint/SuperGlue find ~10× fewer keypoints and the recovered poses come out tens of degrees off). highres_for_matching/ holds the 3072²/3200² image the annotations were actually computed from, so preprocessing can be reproduced:

mkdir -p data/coffee/raw
cp DuplicateSingleImage/train_split/coffee/highres_for_matching/* data/coffee/raw/
python preprocess/run.py --instance_dir data/coffee --crop_size 1984 --fix_focal

--crop_size must fit the largest instance bounding box in the high-res image, with slack for rotation — per object: airplane 1216, box 1536, cake 1792, cash 1472, cheese 1536, cleaner 1536, clock 960, coffee 1984, cola 1472, fire 1792, gitar 2112, potato 1280, sign 1472, tin 1472, yogurt 1728.

Notes:

• It exists under train_split only; eval_split would be a byte-for-byte duplicate.
• It contains only the two files stage 0 reads, not high-res GT normal/albedo/roughness maps.
• These are 8-bit PNGs rather than the HDR .exr the synthetic objects were rendered to; the tonemapping differs slightly from train/000_rgb.exr, which does not affect keypoint matching.
• potato's segmentation is a 4× nearest-neighbour upsample of the 800 px one (no high-res segmentation exists upstream); every other object's is native resolution.
• A re-run will not match the released poses bit-for-bit — COLMAP's gauge is arbitrary. Compare gauge-invariantly (relative rotations) or via the trainer's dr/dt. For coffee a re-run scores dr = 0.54°, dt = 0.013 against the Blender GT versus dr = 0.59°, dt = 0.011 for the released annotation.
• 117 MB in total. Skip it with --exclude "train_split/*/highres_for_matching/*".

Citation

@inproceedings{cheng2023structure,
  title={Structure from Duplicates: Neural Inverse Graphics from a Pile of Objects},
  author={Cheng, Tianhang and Ma, Wei-Chiu and Guan, Kaiyu and Torralba, Antonio and Wang, Shenlong},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}