yoavram-lab/AnimalCLEF2026-sub231

AnimalCLEF26 — sub#231 artefacts (0.81655)

Model checkpoints and precomputed similarity / high-confidence-match matrices for AnimalCLEF 2026 competition submission sub#231 = 0.81655 (2nd place era best; AnimalCLEF26 reID, 4 datasets: Salamander, Lynx, SeaTurtle, Texas). See description in manuscript:

Yesharim M, Ram Y. 2026. Global Embeddings, Local Feature Matching, and Adaptive Validation Overfitting in Animal Re-Identification. CLEF 2026 Working Notes. https://ceur-ws.org/Vol-4283/paper372.pdf

These are the heavy artefacts of the standalone sub231/ reproduction bundle. All of them are regenerable from raw data with the scripts in that bundle (see its README); they are hosted here only because they are ~1.2 GB. Download them back into the bundle with scripts/download_artefacts.py.

Contents

file	kind	dataset	size	sha256	description
resources/sal_head_sp_v2_hc.npy	matrix	Salamander	17.3 MB	c446a6df6588…	head SuperPoint all-pairs HC (v2 crops)
resources/sal_body_sp_profile_rot180_hc.npy	matrix	Salamander	17.3 MB	94ec4c8aa998…	body SuperPoint rot180 TTA HC
resources/sal_convnext_small_selfsup_simv1.npy	matrix	Salamander	17.3 MB	22836eacd0fc…	self-trained ConvNeXt cross-orient sim
resources/sal_hc_sp_allpairs.npy	matrix	Salamander	17.3 MB	c281eb8719ce…	body SuperPoint all-pairs base HC
resources/sal_head_sp_hc.npy	matrix	Salamander	17.3 MB	444e76b0a68b…	head SuperPoint all-pairs HC (v1 crops)
resources/czechlynx_v5b_ft_sim_all.npy	matrix	Lynx	60.9 MB	25efbad524ea…	v5b MiewID similarity
resources/lynx_dinov2_vitb_sim.npy	matrix	Lynx	121.9 MB	4ef94d26727f…	DINOv2-ViT-B 25ep similarity
resources/turtle_ext_ft_sim_all.npy	matrix	SeaTurtle	340.7 MB	a6584c1b64b0…	MiewID FT similarity
resources/texas_aliked_hc.npy	matrix	Texas	0.3 MB	8d5ec5f55d01…	ALIKED+LightGlue+RANSAC HC
resources/weights/miewid_czechlynx_v5b_finetuned.pt	checkpoint	Lynx	211.3 MB	a6fb4e230d04…	v5b MiewID checkpoint
resources/miewid_czechlynx_finetuned.pt	checkpoint	Lynx	210.6 MB	eb8b4dd46cd7…	v1 MiewID checkpoint, fine-tuned on CzechLynx only (not used by sub#231; see Provenance)
resources/weights/miewid_turtle_ext_finetuned.pt	checkpoint	SeaTurtle	210.6 MB	0f02d63a01e0…	MiewID FT checkpoint
resources/miewid_sal_finetuned.pt	checkpoint	Salamander	210.6 MB	7d2b48b59c64…	optional MiewID warm-start (unused by sub#231 convnext path)
submission_sub231_CNself055.csv	reference	all	0.1 MB	9f41ba25a09b…	sub#231 submission CSV of record

Provenance

• Lynx uses two fine-tuned models: v5b (MiewID + ArcFace on CzechLynx+comp, finetune_czechlynx.py) and DINOv2-ViT-B 25ep (lynx_generic_selftrain.py, vit_base_patch14_dinov2.lvd142m, job 13551765), blended 0.40/0.60.
• Lynx, v1 checkpoint (resources/miewid_czechlynx_finetuned.pt, added Oct 2026): MiewID-msv3 + 512-d projection head + ArcFace, fine-tuned with finetune_czechlynx.py on the training part of the CzechLynx time-based open-world split only, without the competition labels (unlike v5b). Because the AnimalCLEF lynx images come from CzechLynx, some of the same animals or photos may be in its training data. Not used by sub#231 and not in artefacts_manifest.json. It is the embedding model of the generative models of similarity matrices in yoavram-lab/mrsbi (private; branch gen-miewid, genmodels/README.md).
• SeaTurtle: MiewID fine-tuned on SeaTurtleIDHeads + ZindiTurtleRecall (finetune_turtles.py).
• Salamander: self-trained ConvNeXt-Small cross-orientation bridge (sal_selfsup_crossorient.py) + SuperPoint all-pairs HC signals.
• Texas: ALIKED+LightGlue+RANSAC high-confidence matches.

License / use

Derived from AnimalCLEF26 competition data, the CzechLynx dataset (CC BY-NC-SA 4.0), and from MiewID (conservationxlabs) and DINOv2 (Meta). Redistribution subject to those upstream licenses and the competition rules — verify before any non-research use.

Integrity: each file's sha256 is pinned in artefacts_manifest.json in the bundle.