Update README.md

# YOLO26 — Person Detection Fine-tuned on PRW

Fine-tuned variants of **YOLO26** (Small/ Large) for person detection, trained with a two-stage pipeline: intermediate pre-training on [CrowdHuman](https://www.crowdhuman.org/) followed by fine-tuning on [PRW (Person Re-identification in the Wild)](https://github.com/liangzheng06/PRW-baseline).

---


## Training Pipeline

The pipeline consists of three phases:

### Phase 0 - COCO Zero-shot Baseline
YOLO26 weights pre-trained on COCO are evaluated directly on PRW without any fine-tuning, establishing a zero-shot reference point.

### Phase 1 - CrowdHuman Intermediate Pre-training
The COCO-pretrained model is fine-tuned on CrowdHuman (full-body `fbox` annotations, person class only, `ignore=1` instances excluded) to adapt the backbone to pedestrian/crowd scenarios before seeing PRW.

**CrowdHuman training settings:**
- Epochs: 30 | Optimizer: AdamW | LR: 5e-4 → 5e-6 (cosine)
- Warmup: 3 epochs | Early stopping patience: 10
- Input size: 640×640

### Phase 2 - Fine-tuning Strategy Ablation (Small only)
Two fine-tuning strategies are compared starting from the best Phase-1 weights:

| Strategy | Frozen layers | LR |
|---|---|---|
| Full FT | None | 1e-4 |
| Partial FT | Backbone layers 0–9 | 3e-4 |

The winning strategy is then applied to Large.

**PRW fine-tuning settings:**
- Epochs: 20| Optimizer: AdamW | Warmup: 5 epochs
- Early stopping patience: 15
- Augmentation: mosaic=1.0, mixup=0.1, HSV jitter, horizontal flip

### Phase 3 - Scale-up to Medium & Large
The best Phase-2 strategy is applied to YOLO26-Large. If CrowdHuman pre-training was beneficial in Phase 1, Large also receive CrowdHuman intermediate pre-training before PRW fine-tuning (same hyperparameters as Phase 1).

---

## Datasets

| Dataset | Split | Images | Boxes | Use |
|---|---|---|---|---|
| [CrowdHuman](https://www.crowdhuman.org/) | train | ~15 000 | ~470 000 | Intermediate pre-training |
| [CrowdHuman](https://www.crowdhuman.org/) | val | ~4 370 | ~137 000 | Pre-training validation |
| [PRW](https://github.com/liangzheng06/PRW-baseline) | train | ~5 700 | ~18 000 | Fine-tuning |
| [PRW](https://github.com/liangzheng06/PRW-baseline) | test | ~6 100 | ~25 000 | Evaluation |

---

README.md

@@ -1,3 +1,13 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+language:
+- en
+- it
+metrics:
+- precision
+- recall
+pipeline_tag: object-detection
+tags:
+- code
+- legal
+---