3D-LENS

Checkpoints of 3D-LENS: A 3D Lifting-based Elevated Novel-view Synthesis method for Single-View Aerial-Ground Re-Identification (ECCV 2026): paper, arXiv:2604.26520, poster. Code, training and evaluation are at https://github.com/TurtleSmoke/3D-LENS.

Checkpoints

Each checkpoint is trained on one camera type of AG-ReID (agreid_*) or AG-ReID.v2 (agreid_v2_*), given by its name. The two AG-ReID.v2 aerial checkpoints are evaluated against CCTV and wearable images respectively.

Checkpoint	Query → Gallery	mAP	Rank-1
agreid_aerial.pth	A → G	49.78	62.84
	G → A	49.44	57.59
agreid_ground.pth	A → G	53.26	65.17
	G → A	55.05	67.36
agreid_v2_aerial_cctv.pth	C → A	37.84	47.10
	A → C	42.77	55.69
agreid_v2_cctv.pth	C → A	47.21	58.53
	A → C	41.36	50.76
agreid_v2_aerial_wearable.pth	W → A	28.61	35.77
	A → W	34.26	47.13
agreid_v2_wearable.pth	W → A	38.27	46.24
	A → W	38.93	47.89

A = aerial, G = ground, C = CCTV, W = wearable. The checkpoints are stored in fp16, so they can differ slightly from the paper.

Usage

git clone https://github.com/TurtleSmoke/3D-LENS && cd 3D-LENS && uv sync && source .venv/bin/activate
hf download TurtleSmoke/3D-LENS --include "*.pth" --local-dir checkpoints
bash scripts/eval.sh agreid_aerial $DATA checkpoints/agreid_aerial.pth

License

The weights are released for non-commercial research under CC BY-NC 4.0. They were trained on AG-ReID and AG-ReID.v2 and on synthetic views generated with Hunyuan3D-2 and StyleID, whose respective terms also apply. The code is under the CeCILL-B license.

Citation

@inproceedings{grolleau2026lens,
    title = {3D-LENS: A 3D Lifting-based Elevated Novel-view Synthesis method for Single-View Aerial-Ground Re-Identification},
    author = {Grolleau, William and Sabourin, Astrid and Lapouge, Guillaume and Achard, Catherine},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year = {2026}
}