
N-Benchmark Top-1	86.6% (110/127)
Top-3 / Top-5 / Top-10	96.9% / 97.6% / 100.0%
Avg Rank / Worst Rank	1.30 / 9
Backbone	DINOv2 ViT-B/14 (86.6M params, timm 518×518)
Loss	Triplet (margin=0.3) + Hard Mining
Sampler	AreaAwareSampler (area_ratio=0.75)
Training	4-phase progressive unfreezing, 56 epochs, ~7.2h
Embedding	1280-d, L2-normalized
Files	`e3_01b_dinov2_vitb_best/`

Most Efficient — DINOv3 ViT-S+/16


N-Benchmark Top-1	81.1% (103/127)
Top-3 / Top-5 / Top-10	92.1% / 95.3% / 99.2%
Avg Rank	1.61
Backbone	DINOv3 ViT-S+/16 (~22M params, timm 512×512)
Loss	Triplet (margin=0.3) + Hard Mining
Sampler	MPerClassSampler (m=2)
Training	4-phase progressive unfreezing, 63 epochs, ~2.0h
Embedding	768-d, L2-normalized
Files	`dinov3_vitsplus_efficient/`

Comparison

Metric	Best Accuracy	Most Efficient	Difference
Top-1	86.6%	81.1%	-5.5%
Parameters	~86.6M	~22M	-75%
Model size (FP32)	338 MB	112 MB	-67%
Model size (FP16)	169 MB	56 MB	-67%
Training time	~7.2h	~2.0h	-72%
Inference tokens	1369 (patch14)	1024 (patch16)	-25%

Quantized Variants

Both models are available in FP16 with zero accuracy loss — recommended for GPU deployment.

Model	Format	Size	Top-1	Avg Rank
DINOv2 ViT-B	FP32	338 MB	86.6%	1.30
DINOv2 ViT-B	FP16	169 MB	86.6%	1.30
DINOv3 ViT-S+	FP32	112 MB	81.1%	1.61
DINOv3 ViT-S+	FP16	56 MB	81.1%	1.61

FP16 inference requires CUDA. For CPU inference, use FP32.

Quick Start

bash

# Install dependencies (standalone, no coral_reid needed)
uv sync

# Extract features from a single image
uv run python extract_features.py \
    --model e3_01b_dinov2_vitb_best/best_model_20260308_110634.pt \
    --input /path/to/image.jpg

# Extract features from a directory
uv run python extract_features.py \
    --model e3_01b_dinov2_vitb_best/best_model_20260308_110634.pt \
    --input /path/to/images/ \
    --output features.h5

N-Benchmark Per-Area Results

E3-01b DINOv2 ViT-B/14 (Best)

Area	Queries	Top-1	Top-3	Top-5	Avg Rank
37	32	93.8%	96.9%	96.9%	1.28
38	31	80.6%	100.0%	100.0%	1.19
39	27	85.2%	92.6%	96.3%	1.44
40	37	86.5%	97.3%	97.3%	1.30
Overall	127	86.6%	96.9%	97.6%	1.30

DINOv3 ViT-S+/16 (Efficient)

Area	Queries	Top-1	Top-3	Top-5	Avg Rank
37	32	81.2%	93.8%	96.9%	1.56
38	31	77.4%	90.3%	93.5%	1.90
39	27	85.2%	92.6%	96.3%	1.37
40	37	81.1%	91.9%	94.6%	1.57
Overall	127	81.1%	92.1%	95.3%	1.61

Full Model History

Model Comparison Table

Model Name	Arch	Backbone	Loss	Mining	Same Area Neg	Image	Test Acc	Test Loss	Val Loss	N-Bench Avg	A37	A38	A39	A40	Time
Pre-trained	-	DINOv2-B/14	-	-	-	bbox	-	-	-	29.48%	28.12%	35.48%	29.63%	24.32%	-
20250812_152526	old	DINOv2-B/14	Triplet	pre-composed	❌	bbox	92.6%	0.1659	-	48.25%	50.00%	51.61%	48.15%	43.24%	~16h
20251007_133126	old	DINOv2-B/14	Triplet	pre-composed	✅	bbox	88.8%	0.2523	-	39.32%	46.88%	41.94%	33.33%	35.14%	~16h
20251008_094017	old	DINOv2-B/14	Triplet	pre-composed	✅	bbox	90.4%	0.1636	-	40.19%	37.50%	48.39%	37.04%	37.84%	~16h
20251014_183603	old	DINOv2-B/14	Triplet	pre-composed	❌	bbox	92.8%	0.1012	-	40.97%	37.50%	38.71%	44.44%	43.24%	~16h
Pre-trained	-	DINOv2-B/14	-	-	-	whole	-	-	-	50.88%	34.38%	54.84%	62.96%	51.35%	-
20251015_165008	old	DINOv2-B/14	Triplet	pre-composed	✅	whole	92.7%	0.1330	0.1006	64.43%	62.50%	61.29%	55.56%	78.38%	~16h
20251016_133229	old	DINOv2-B/14	Triplet	pre-composed	❌	whole	97.9%	0.0429	-	63.31%	56.25%	58.06%	74.07%	64.86%	~16h
20260308_110634	new	DINOv2-B/14 (timm 518)	Triplet	dynamic (PML)	AreaAware 0.75	whole	-	-	0.1604	86.6%	93.8%	80.6%	85.2%	86.5%	~7.2h
20260306_233824	new	DINOv3-S+/16 (timm 512)	Triplet	dynamic (PML)	❌	whole	-	-	0.1604	81.1%	81.2%	77.4%	85.2%	81.1%	~2.0h

Column Descriptions

Column	Description
Arch	`old` = old_repo implementation, `new` = refactored modular architecture
Backbone	Feature extractor (DINOv2-B/14, DINOv3-S+/16, etc.)
Loss	Loss function (Triplet, ArcFace, CosFace, Circle, Contrastive, etc.)
Mining	Sample mining: `pre-composed` = fixed triplets, `dynamic (PML)` = MPerClassSampler
Same Area Neg	Whether negatives restricted to same geographic area (`AreaAware 0.75` = 75% same area)
Image	`bbox` = EXIF bounding box crop, `whole` = full image
Test Acc	Test set accuracy (old arch only, measures pos_dist < neg_dist)
Val Loss	Best validation loss during training
N-Bench Avg	N-Benchmark Top-1 accuracy averaged across areas 37-40

Architecture Differences

Feature	Old Architecture	New Architecture
Dataset Output	`(anchor, pos, neg)` - 3 images	`(image, label)` - 1 image
Triplet Formation	Pre-composed before training	Dynamic mining per batch
Batch Sampler	Random	MPerClassSampler (m=2)
Loss Function	Custom TripletLossWithMining	PML TripletMarginLoss
Samples per Epoch	~50,000 triplets x 3 images	~4,000 images
Training Speed	~23 min/epoch	~1.5 min/epoch
Same Area Negatives	Implemented	Implemented (AreaAwareSampler)

N-Benchmark (Nearest Benchmark): Top-1 accuracy rate of identifying the correct coral when comparing specimens in areas 37-40 across 2022 and 2023.

Project Structure

coral_models/
├── pyproject.toml                        # uv environment (standalone)
├── extract_features.py                   # Feature extraction script
├── e3_01b_dinov2_vitb_best/              # Best accuracy model (86.6%)
│   ├── best_model_20260308_110634.pt      # FP32 (338 MB)
│   ├── best_model_20260308_110634_fp16.pt # FP16 (169 MB)
│   ├── final_model_20260308_110634.pt
│   ├── e3_01b_same_area_neg_075.yaml
│   ├── README.md
│   └── README(zh-tw).md
├── dinov3_vitsplus_efficient/            # Most efficient model (81.1%)
│   ├── best_model_20260306_233824.pt      # FP32 (112 MB)
│   ├── best_model_20260306_233824_fp16.pt # FP16 (56 MB)
│   ├── final_model_20260306_233824.pt
│   ├── dinov3_vitsplus_tune_02_p2lr5_4ph.yaml
│   ├── README.md
│   └── README(zh-tw).md
└── legacy/                              # Old architecture models (torch.hub, 224×224)
    ├── dinov2_coral_best_model_20251015_165008.pt      # 64.43% (old best, same area neg, whole image)
    └── dinov2_coral_best_model_20251016_133229.pt      # 63.31% (no same area neg, whole image)

License

This project is licensed under GPL-3.0.

Based on DINOv2 and DINOv3 by Meta Platforms, Inc. (Apache License 2.0).

YuC13600/coral_models

Coral Re-Identification Models

Best Models

Best Accuracy — E3-01b DINOv2 ViT-B/14