Balanced Dog Pose Estimation Dataset (B-DoPED)

Companion dataset to the ECCV 2026 paper "Every Dog Has Its Day, Probably" (see Citation).

B-DoPED is a synthetic, multi-view dataset of articulated 3D dogs built on the D-SMAL model — the dog-specialized variant of SMAL introduced by BITE (model type 39dogs_norm). It ships three reusable parameter libraries (pose / shape / texture), a standalone rendering script, and a large set of pre-rendered multi-view outputs.

Contents

library/
  poses/poses.npz        # pose_6d (35000, 34, 6) float32  — 35k articulated poses (6D joint rotations)
  shapes/shapes.npz      # beta (500,30), betas_limbs (500,9), pose_6d (500,34,6), logscale_part_list (9,)
  textures/              # 1000 flat PNGs (2048²) + one shared UV unwrap
    texture_00000.png … texture_00999.png
    uv_atlas.pth         # shared UV unwrap {vmapping (5187,), faces (7774,3), uvs (5187,2)}
    material.mtl
scripts/                 # standalone renderer + setup
  setup.sh  render_smal_multiview.py  render_utils.py  smal_utils.py  shard_utils.py  load_examples.py
docs/SMAL_SETUP.md
outputs/                 # pre-rendered multi-view results, packed as tar shards
  shard_00000.tar … shard_00019.tar   # 20 shards × 1750 poses; in-tar: pose_NNNNNN/{rgb,seg,npz}/VV.{png,npz}
  shards_index.csv       # per-shard pose range, file/byte counts, sha256
environment.yml  requirements.txt  LICENSE  .gitattributes

Conventions. A pose is pose_6d (34, 6) — 6D rotations for 34 joints. Row k of poses.npz corresponds to the directory pose_{k:06d}/ in the render shards. Each pose was rendered from 60 camera views (4 azimuths × 5 elevations × 3 rolls) at 256×256. For each view a random shape and texture were sampled; the shape-specific ear pose (shape pose_6d, joints 32/33) is blended into the motion pose.

Per-view NPZ schema (outputs/.../npz/VV.npz)

key	shape	meaning
camera/scale, camera/tx, camera/ty	scalar	weak-perspective camera
smal/beta	(1, 30)	D-SMAL shape
smal/betas_limbs	(1, 9)	limb log-scales (logscale_part_list)
smal/pose_6d	(1, 34, 6)	joint pose (6D), ear-blended
smal/orient_6d	(1, 6)	camera/global orientation
smal/trans	(1, 3)	translation
smal/vert_off_compact	(1, 5901)	per-vertex offsets (compact)
smal/keyp_3d_all	(1, 47, 3)	3D keypoints
smal/keyp_2d_all	(1, 47, 2)	2D keypoints, normalized to [-1, 1]
smal/keyp_conf	str	keypoint configuration (all)
smal/logscale_part_list	(9,)	limb part names
smal/smal_model_type	str	39dogs_norm

Quickstart

# 1. (optional) only the library + scripts, not the 92 GB of renders:
hf download 1Konny/B-DoPED --repo-type dataset --local-dir ./pups --exclude "outputs/*"

# 2. environment + SMAL/BITE dependency
conda env create -f environment.yml && conda activate eccv26_bdoped
bash scripts/setup.sh                 # fetches BITE code + D-SMAL weights (see docs/SMAL_SETUP.md)

# 3. peek at the data
python scripts/load_examples.py

Use the libraries directly

import numpy as np
poses  = np.load("library/poses/poses.npz")["pose_6d"]          # (35000, 34, 6)
shapes = np.load("library/shapes/shapes.npz", allow_pickle=True)
betas, limbs = shapes["beta"], shapes["betas_limbs"]            # (500,30), (500,9)

Read the rendered outputs (stream a shard, no unpack)

import io, tarfile, numpy as np
with tarfile.open("outputs/shard_00000.tar") as tar:
    npz = tar.getmember("pose_000000/npz/00.npz")
    d = np.load(io.BytesIO(tar.extractfile(npz).read()), allow_pickle=True)
    print(sorted(d.files))

Render your own

python scripts/render_smal_multiview.py \
  --pose_npz library/poses/poses.npz \
  --shape_npz library/shapes/shapes.npz \
  --texture_dir library/textures \
  --output_root ./my_renders --bite_root ./bite_gradio-hf \
  --device cuda --resolution 256 --pose_indices_file <(seq 0 9)
# tail_drop_prob 0.1 reproduces the released generation; --shard packs outputs into tar shards.

Partial / selective download

It is a single repository, but the renders are optional:

hf download 1Konny/B-DoPED --repo-type dataset --exclude "outputs/*"          # library + scripts only
# or in Python:
from huggingface_hub import snapshot_download
snapshot_download("1Konny/B-DoPED", repo_type="dataset", ignore_patterns=["outputs/*"])
# git alternative (skip large LFS blobs, then fetch what you want):
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/1Konny/B-DoPED

License

Released for non-commercial academic research only under the Research-Only Terms of Use. The authors claim no ownership of the upstream models/assets the dataset derives from and grant no rights beyond those allowed upstream — you are responsible for complying with all applicable upstream licenses (e.g. the D-SMAL / BITE model; its weights are not redistributed here and must be obtained separately, see docs/SMAL_SETUP.md). The Dataset is provided as is, without warranty.

Citation

This dataset accompanies the following paper. If you use it, please cite:

@inproceedings{choi2026everydog,
  title     = {Every Dog Has Its Day, Probably: A Balanced Synthetic Benchmark and
               Probabilistic Modeling for 3D Dog Pose Estimation},
  author    = {Choi, Joo Young and Lee, Wonkwang and Seon, Ju-hyeong and Kim, Gunhee},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026},
}