Upload 14 files

README.md

@@ -8,36 +8,97 @@ tags:
   - yolo11s
 library_name: ultralytics
 license: apache-2.0
+datasets:
+  - traffic-sign-detection-vietnam
 ---
 
-# 🚦 Traffic Sign Detection — Vietnam
+# 🚦 Traffic Sign Detection — Vietnam (YOLO11s)
 
-YOLO11s model trained on the Vietnam Traffic Sign Detection dataset (~10k images, 82 classes).
+[![YOLO](https://img.shields.io/badge/YOLO-v11-00c2cb?style=for-the-badge&logo=ultralytics&logoColor=white)](https://github.com/ultralytics/ultralytics)
+[![Dataset](https://img.shields.io/badge/Dataset-Hugging%20Face-orange?style=for-the-badge&logo=huggingface)](https://huggingface.co/datasets/star092304/Traffic-sign-detection-VietNam)
 
-## Metrics (test set)
+YOLO11s model trained on the Vietnam Traffic Sign Detection dataset.
+
+| Property | Value |
+|----------|-------|
+| Model | YOLO11s |
+| Classes | 82 Vietnamese traffic signs |
+| Image size | 640×640 |
+| Framework | Ultralytics |
+| Dataset Source | [Hugging Face](https://huggingface.co/datasets/star092304/Traffic-sign-detection-VietNam) |
+
+## Evaluation Results
+
+Below is the summary of the evaluation results from `evalution/summary.json`:
 
 | Metric | Value |
 |--------|-------|
-| mAP@0.5 | 0.9806 |
-| mAP@0.5:0.95 | 0.8357 |
-| Precision | 0.9642 |
-| Recall | 0.9615 |
-| FPS (T4 GPU) | 61.5 |
+| **Model** | `yolo11s.pt` |
+| **Epochs Trained** | 50 |
+| **Number of Classes** | 82 |
+| **Device** | GPU |
+| **Precision** | 96.42% (`0.9642`) |
+| **Recall** | 96.15% (`0.9615`) |
+| **mAP50** | 98.06% (`0.9806`) |
+| **mAP75** | 93.37% (`0.9337`) |
+| **mAP50-95** | 83.57% (`0.8357`) |
+| **FPS** | 61.5 |
+| **Mean Latency** | 16.25 ms |
+| **p50 Latency** | 15.03 ms |
+| **p95 Latency** | 22.59 ms |
+| **Min Latency** | 13.44 ms |
+| **Max Latency** | 23.24 ms |
+
+## Visualizations
+
+### Training Curves
+![Training Curves](training/training_curves.png)
+
+### Results
+![Results](training/results.png)
+
+### Random Predictions (Inference)
+![Random Predictions](inference/random_predictions.png)
 
-## Usage
+## Files
+
+| File | Description |
+|------|-------------|
+| `best.pt` | PyTorch weights (main model) |
+| `best.onnx` | ONNX export (CPU/edge deploy) |
+| `data.yaml` | Dataset config with class names |
+| `config.json` | Training hyperparameters |
+| `metrics.json` | Test-set evaluation results |
+| `benchmark.json` | FPS / latency results |
+| `summary.json` | All metrics combined |
+
+## Quick Start
 
 ```python
 from ultralytics import YOLO
-model = YOLO("best.pt")   # or best.onnx for ONNX
-results = model("image.jpg")
+
+# PyTorch
+model = YOLO("best.pt")
+results = model("image.jpg", conf=0.25)
 results[0].show()
+
+# ONNX (faster on CPU)
+model_onnx = YOLO("best.onnx")
+results = model_onnx("image.jpg")
+```
+
+## CLI Inference
+
+```bash
+python infer.py --source image.jpg
+python infer.py --source video.mp4 --save
+python infer.py --source 0 --show          # webcam
 ```
 
-## Classes
-82 traffic sign categories for Vietnam road conditions.
+## Training Details
 
-## Training Config
-- Model: yolo11s.pt
-- Epochs: 50 (early stopping, patience=20)
-- Image size: 640
-- Batch: 16
+- **Base model**: yolo11s.pt (pretrained COCO)
+- **Early stopping**: patience=20
+- **Dataset cache**: enabled (faster I/O)
+- **Seed**: 42 (reproducible)
+- **Optimizer**: auto (AdamW)

evalution/benchmark.json

@@ -0,0 +1,9 @@
+{
+  "mean_latency_ms": 16.25,
+  "p50_latency_ms": 15.03,
+  "p95_latency_ms": 22.59,
+  "min_latency_ms": 13.44,
+  "max_latency_ms": 23.24,
+  "fps": 61.5,
+  "device": "GPU"
+}

evalution/metrics.json

@@ -0,0 +1,7 @@
+{
+  "mAP50": 0.9805747962455071,
+  "mAP50_95": 0.8356933876101033,
+  "mAP75": 0.9337388412187322,
+  "precision": 0.964222944759273,
+  "recall": 0.9614566153284894
+}

evalution/per_class_metrics_test.csv

@@ -0,0 +1,83 @@
+class,precision,recall,ap50
+Low Clearance,0.9900201157335413,1.0,0.995
+No Trucks and Bus,1.0,0.9850546469510434,0.995
+No Left Turn,0.9865452139078744,1.0,0.995
+No Horns,0.9576904085232468,1.0,0.995
+Danger,0.9879077354665579,1.0,0.995
+No Cars,0.9850916160496327,1.0,0.995
+Double curve first to right,0.9859768977459309,1.0,0.995
+No Moto,0.9507722372175216,1.0,0.995
+Road with Surveillance Camera,0.8743140903719894,1.0,0.995
+Obstacle on the Road,1.0,0.905884424847886,0.995
+Lane Allocation,0.9908375001914379,1.0,0.995
+Sharp Right Turn,0.9193752700402034,1.0,0.995
+Intersection with a Minor Road,1.0,0.9916006767002455,0.995
+Intersection with Equal Roads,0.9978528486386194,1.0,0.995
+Intersection with a Priority Road,0.9712350750376608,1.0,0.995
+No Straight and Right Turn,0.9638621972390922,1.0,0.995
+Traffic light ahead,0.8869606369280012,1.0,0.995
+No U-Turn and No Left Turn,0.9789871401989048,1.0,0.995
+End of 50km/h speed limit,0.9802774078441459,1.0,0.995
+sparsely populated area,0.9729635140617088,1.0,0.995
+Slippery Road,0.955235716202147,1.0,0.995
+Uneven road,0.9092611371922793,1.0,0.995
+Steep ascent,0.9794949177980316,1.0,0.995
+No left turn for cars,0.9837367908670516,1.0,0.995
+No Parking on Even Days,0.9637730802776739,1.0,0.995
+No Left or Right Turn,0.9883674415684336,1.0,0.995
+Pedestrian Lane,0.947527671534424,1.0,0.995
+No Motobike Left Turn,0.953364060077053,1.0,0.995
+No bus,0.9527315374075699,1.0,0.995
+No Overtaking,0.9679110133888323,1.0,0.995
+Children Crossing,0.988352038941514,1.0,0.995
+Narrow road both sides,0.946315305789602,1.0,0.995
+Narrow Road Right Side,0.9829978478935989,1.0,0.995
+Narrow Road Left Side,0.9607601156068228,1.0,0.995
+Keep left,0.985278824607956,1.0,0.995
+End of all prohibition,0.9740146516172447,1.0,0.995
+No U-Turn for Cars,0.8858557525212286,1.0,0.995
+Level Crossing with Barriers,1.0,0.9313327581927153,0.995
+No U-Turn and Left Turn for Cars,0.9302647395645418,1.0,0.995
+Narrow bridge,0.9758132364033345,1.0,0.995
+Stop,0.9725653895150188,1.0,0.995
+U-Turn Area,0.9944662320261956,1.0,0.995
+Speed limit 100km/h,0.9745238351976454,1.0,0.995
+Speed limit 110km/h,0.9741028723376567,1.0,0.995
+Speed limit 120km/h,0.9205756957498799,1.0,0.995
+Speed limit 90km/h,1.0,0.8172760587020905,0.995
+Roundabout,0.9880233147700539,1.0,0.995
+Turn Right,0.9395931640615742,1.0,0.995
+Turn Left,0.983730793325145,1.0,0.995
+Road Work Ahead,0.9496755252994297,1.0,0.995
+Hospital,0.9807015621273367,1.0,0.995
+Parking,0.9535981311686142,1.0,0.995
+No Parking on Odd Days,0.976600523189776,1.0,0.995
+No Stopping & No Parking,0.9919736228102467,0.9868421052631579,0.9948051948051947
+No Two or Three-wheeled Vehicles,1.0,0.9726513082883271,0.9942000000000001
+Speed limit 80km/h,0.9488388275467288,1.0,0.9940243902439024
+No Right Turn,0.9700964150830149,0.9411764705882353,0.9930065359477124
+No Trucks,0.9742164412587491,0.9615384615384616,0.9918838288003374
+Height Limit,0.9731698070359129,0.9642857142857143,0.9904545454545456
+Accident area,0.933089420398414,1.0,0.9892105263157894
+Speed limit 10km/h,0.9477264927429047,0.9,0.985909090909091
+Pedestrian Crossing,1.0,0.9450610873861024,0.985
+No Parking,0.9899759476073186,0.978494623655914,0.9846842105263157
+Slow Down,0.969269324416295,0.9841269841269841,0.9837301587301587
+Speed limit 40km/h,1.0,0.8845621142739079,0.9819442967362624
+Sharp Left Turn,0.9411159594628086,0.9259259259259259,0.9817369727047145
+Speed limit 20km/h,1.0,0.8741236122229261,0.9811538461538463
+Speed limit 60km/h,0.9708657751325349,0.9387480064919294,0.9811391248697522
+Speed limit 70km/h,0.9340237201000852,0.8332760484717904,0.9784173669467788
+Dual carriageway,0.9491643245781851,0.983286491563703,0.9770000000000001
+No Entry,0.9833202557287564,0.8544015257941685,0.965110483403822
+No U-Turn,0.9638881708382964,0.967741935483871,0.965
+Bus Stop,0.9635511195931211,0.9444516575857006,0.9535714285714283
+Turn Right Only,0.9698898062735045,0.7936507936507936,0.9416035081665655
+One way street,0.9286031948005872,0.92,0.9414367816091955
+Green Light,0.9375582592006058,0.7766495726426715,0.924665723373578
+Speed Bump,0.9733887965476249,0.8666666666666667,0.916578947368421
+Speed limit 30km/h,0.9330278899269068,0.9166666666666666,0.915
+Residential area,0.9703748843314238,0.9166666666666666,0.915
+No U-Turn and No Right Turn,0.9085002818929591,0.9090909090909091,0.905
+Speed limit 50km/h,1.0,0.6954774938108672,0.8478196930946291
+Red Light,0.820769909729665,0.5727310494000819,0.7130466373995354

evalution/summary.json

@@ -0,0 +1,17 @@
+{
+  "mAP50": 0.9805747962455071,
+  "mAP50_95": 0.8356933876101033,
+  "mAP75": 0.9337388412187322,
+  "precision": 0.964222944759273,
+  "recall": 0.9614566153284894,
+  "mean_latency_ms": 16.25,
+  "p50_latency_ms": 15.03,
+  "p95_latency_ms": 22.59,
+  "min_latency_ms": 13.44,
+  "max_latency_ms": 23.24,
+  "fps": 61.5,
+  "device": "GPU",
+  "num_classes": 82,
+  "model": "yolo11s.pt",
+  "epochs_trained": 50
+}

inference/infer_local.py

@@ -0,0 +1,129 @@
+"""
+Traffic Sign Detection — Inference CLI (Local & Hugging Face integrated)
+Usage:
+    python infer.py --source image.jpg
+    python infer.py --source test.mp4 --save
+    python infer.py --source 0               # webcam
+    python infer.py --source images/ --save
+
+    # Custom local weights path:
+    python infer.py --source image.jpg --weights outputs/best.pt
+"""
+
+import argparse
+import time
+from pathlib import Path
+
+import torch
+import numpy as np
+from ultralytics import YOLO
+from huggingface_hub import hf_hub_download
+
+
+def parse_args():
+    p = argparse.ArgumentParser(description="Traffic Sign Inference CLI")
+    p.add_argument(
+        "--source", required=True, help="Image/video path, folder, or 0 for webcam"
+    )
+    p.add_argument(
+        "--weights",
+        default="best.pt",
+        help="Path to model weights (.pt or .onnx) or filename on Hugging Face repo",
+    )
+    p.add_argument("--conf", type=float, default=0.25, help="Confidence threshold")
+    p.add_argument("--iou", type=float, default=0.45, help="IoU threshold for NMS")
+    p.add_argument("--imgsz", type=int, default=640, help="Inference image size")
+    p.add_argument("--save", action="store_true", help="Save output images/video")
+    p.add_argument(
+        "--show", action="store_true", help="Display results inside an OpenCV window"
+    )
+    p.add_argument(
+        "--out_dir", default="outputs/infer", help="Output directory for saved results"
+    )
+    return p.parse_args()
+
+
+def main():
+    args = parse_args()
+
+    # 1. Check weights path. If not found locally, download from Hugging Face repository
+    weights_path = Path(args.weights)
+    if not weights_path.exists():
+        repo_id = "star092304/traffic-sign-detection-vietnam-yolo"
+        filename = weights_path.name  # Expecting 'best.pt' or specific weight filename
+        print(f"Local weights '{args.weights}' not found.")
+        print(f"Downloading from Hugging Face repository: {repo_id}...")
+        try:
+            downloaded_path = hf_hub_download(repo_id=repo_id, filename=filename)
+            args.weights = downloaded_path
+            print(f"Successfully downloaded weights to local cache: {downloaded_path}")
+        except Exception as e:
+            print(f"Error downloading from Hugging Face: {e}")
+            print(
+                "Please ensure your internet connection or provide a valid local path to --weights."
+            )
+            return
+
+    # 2. Load the YOLO model
+    print(f"Loading model: {args.weights}")
+    model = YOLO(args.weights)
+
+    # 3. Model GPU Warm-up (highly recommended for local inference speed benchmarks)
+    if torch.cuda.is_available():
+        dummy = np.zeros((args.imgsz, args.imgsz, 3), dtype="uint8")
+        model(dummy, verbose=False)
+        print("GPU Warm-up complete.")
+
+    # 4. Handle webcam source input mapping (convert '0' string to int 0)
+    source_input = args.source
+    if source_input.isdigit():
+        source_input = int(source_input)
+
+    # 5. Run inference pipeline
+    print(f"Running inference on: {args.source}")
+    t0 = time.perf_counter()
+
+    results = model.predict(
+        source=source_input,
+        conf=args.conf,
+        iou=args.iou,
+        imgsz=args.imgsz,
+        save=args.save,
+        show=args.show,
+        project=args.out_dir,
+        name="run",
+        exist_ok=True,
+        verbose=True,
+    )
+
+    elapsed = time.perf_counter() - t0
+
+    # Calculate performance metrics
+    # In case of webcam or video stream, results length corresponds to the total processed frames
+    n = len(results) if isinstance(results, list) else 1
+    fps = n / elapsed if elapsed > 0 else 0
+
+    print(f"\n{'='*40}")
+    print(f"Processed : {n} frame(s)")
+    print(f"Time      : {elapsed:.2f}s")
+    print(f"FPS       : {fps:.1f}")
+    if args.save:
+        print(f"Saved to  : {args.out_dir}/run/")
+
+    # 6. Print detections for a single frame evaluation
+    if n == 1 and hasattr(results[0], "boxes"):
+        boxes = results[0].boxes
+        if boxes is not None and len(boxes) > 0:
+            names = model.names
+            print(f"\nDetections ({len(boxes)}):")
+            for box in boxes:
+                cls = int(box.cls.item())
+                conf = float(box.conf.item())
+                xyxy = box.xyxy[0].cpu().numpy().astype(int)
+                print(f"  [{conf:.2f}] {names[cls]:35s}  box={xyxy}")
+        else:
+            print("\nNo detections found.")
+
+
+if __name__ == "__main__":
+    main()