MOO: Multi-view Oriented Observations

Paper (arXiv:2603.04314) · Code · Project page

Presented at the CVPR 2026 Workshop on Computer Vision for Animal Behavior Tracking and Modeling (CV4Animals), non-archival.

MOO provides 128,000 images of 1,000 synthetic cattle identities rendered from 128 calibrated viewpoints, spanning 360° in azimuth and -25° to 90° in elevation, with depth maps and camera parameters for every image.

Protocol

Identities are split 500/500: even IDs for training, odd IDs for testing. Views 0-63 are side views (elevation < 30°), views 64-127 are top views (elevation > 30°). For each test identity, half of its top views and half of its side views are queries, the rest are gallery. The split definition lives in make_splits.py.

Each config (top, side, topside) has train, query and gallery splits. The paper's evaluation protocols (query → gallery) combine them:

from datasets import load_dataset

train = load_dataset("CEAai/MOO", "topside", split="train")

# Top -> Side
query = load_dataset("CEAai/MOO", "top", split="query")
gallery = load_dataset("CEAai/MOO", "side", split="gallery")
# Side -> Top: ("side", "query") vs ("top", "gallery"); Top-Side -> Top-Side: ("topside", ...) for both

The default download includes depth (~35% of the data). To skip it, stream and select columns; only the requested columns are fetched (without streaming=True, whole files are downloaded and the columns dropped afterwards):

query = load_dataset("CEAai/MOO", "top", split="query", streaming=True, columns=["image", "id", "view_id"])

Use .with_format("numpy") to get depth as a float32 array.

Fields

Field	Type	Description
image	image	512×512 RGB render
depth	float32 512×512	depth map; background is 1e10 and can be used as a mask
id	int32	identity, 0-999
view_id	int32	viewpoint index in the 128-point grid
view	string	coarse direction label (front, back-left, top-right, …)
azimuth, elevation	float64	camera angles in degrees
location	float64[3]	camera position in world coordinates
cam2world	float64 4×4	camera-to-world matrix
K	float64 3×3	camera intrinsics

Citation

@article{grolleau2026moo,
    title = {MOO: A Multi-view Oriented Observations Dataset for Viewpoint Analysis in Cattle Re-Identification},
    author = {Grolleau, William and Chaouch, Achraf and Sabourin, Astrid and Lapouge, Guillaume and Achard, Catherine},
    journal = {arXiv preprint arXiv:2603.04314},
    note = {Presented at the CVPR 2026 Workshop on Computer Vision for Animal Behavior Tracking and Modeling (CV4Animals), non-archival},
    year = {2026},
    doi = {10.48550/arXiv.2603.04314}
}