Fine-tuned YOLOv8n face detection for Bengali video captioning (WiderFace)

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+confusion_matrix.png filter=lfs diff=lfs merge=lfs -text
+results.png filter=lfs diff=lfs merge=lfs -text
+test_inference.png filter=lfs diff=lfs merge=lfs -text
+val_batch0_pred.jpg filter=lfs diff=lfs merge=lfs -text

DATASET_CARD.md

@@ -0,0 +1,57 @@
+---
+language:
+- en
+tags:
+- face-detection
+- yolov8
+- wider-face
+- object-detection
+license: apache-2.0
+task_categories:
+- object-detection
+source_datasets:
+- wider_face
+---
+
+# Dataset Card — WiderFace for YOLO Training
+
+## Source
+
+Kaggle: [lylmsc/wider-face-for-yolo-training](https://www.kaggle.com/datasets/lylmsc/wider-face-for-yolo-training)
+
+## Used To Train
+
+[Tmanna/yolov8-face-bengali-video](https://huggingface.co/Tmanna/yolov8-face-bengali-video)
+
+## Format
+
+YOLO annotation format:
+```
+# Each label .txt file:
+# class_id  cx  cy  width  height  (all normalized 0–1)
+0 0.512 0.340 0.125 0.210
+```
+
+## Class Map
+
+| ID | Class |
+|----|-------|
+| 0  | face  |
+
+## Stats
+
+| Split | Images |
+|-------|--------|
+| Train | ~12,880 |
+| Val   | ~3,226  |
+
+## Original Dataset
+
+WiderFace by Shuo Yang et al. — the standard benchmark for face detection  
+covering extreme scale, pose, occlusion, expression and illumination variation.
+
+## Preprocessing
+
+- Converted to YOLO annotation format  
+- Resized to 640×640 during training  
+- Augmentations: mosaic, horizontal flip, rotation ±10°, HSV jitter, scale, translate

README.md

@@ -0,0 +1,101 @@
+---
+language:
+- bn
+tags:
+- object-detection
+- face-detection
+- yolov8
+- ultralytics
+- bengali-video
+- caption-assignment
+- wider-face
+license: apache-2.0
+datasets:
+- lylmsc/wider-face-for-yolo-training
+metrics:
+- map
+- precision
+- recall
+pipeline_tag: object-detection
+base_model: ultralytics/yolov8n
+---
+
+# YOLOv8n Face Detection — Bengali Video Caption Pipeline
+
+## Model Summary
+
+Fine-tuned **YOLOv8n** for face/person detection in Bengali video frames.  
+Part of a Bengali video auto-captioning system that combines:
+- **This model** → detects visible faces in frames
+- **[Tmanna/whisper-bengali-final](https://huggingface.co/Tmanna/whisper-bengali-final)** → Bengali speech-to-text
+- **Rule-based visibility filter** → shows caption only when face is on screen
+
+## Pipeline Architecture
+```
+Video
+  ↓
+Audio ──→ Whisper (Tmanna/whisper-bengali-final) ──→ Bengali caption text
+  ↓
+Frames ──→ THIS MODEL (YOLOv8 Face Detection)
+  ↓
+Visibility Filter (rule-based: len(boxes) > 0)
+  ↓
+IF face visible → overlay Bengali caption
+ELSE            → skip caption for this frame
+```
+
+## Training Details
+
+| Parameter     | Value |
+|---------------|-------|
+| Base Model    | `yolov8n.pt` (Ultralytics) |
+| Dataset       | [lylmsc/wider-face-for-yolo-training](https://www.kaggle.com/datasets/lylmsc/wider-face-for-yolo-training) |
+| Classes       | 1 (`face`) |
+| Epochs        | 50 |
+| Image Size    | 640 × 640 |
+| Batch Size    | 32 |
+| Hardware      | 2× NVIDIA T4 (Kaggle) |
+| Optimizer     | AdamW + Cosine LR decay |
+| Early Stop    | patience=10 |
+
+## Evaluation Results
+
+| Metric       | Score |
+|--------------|-------|
+| mAP@0.5      | 0.6994 |
+| mAP@0.5:0.95 | 0.3665 |
+| Precision    | 0.8556 |
+| Recall       | 0.6104 |
+
+## Usage
+```python
+from ultralytics import YOLO
+
+model = YOLO("Tmanna/yolov8-face-bengali-video")
+
+# On a video frame (numpy array or image path)
+results = model("frame.jpg", conf=0.25)
+
+faces = results[0].boxes
+if len(faces) > 0:
+    print("Face visible → show Bengali caption")
+else:
+    print("No face → skip caption")
+```
+
+## Visibility Filter (no extra model needed)
+```python
+def should_show_caption(frame_path, face_model, conf=0.25):
+    results = face_model(frame_path, conf=conf, verbose=False)
+    return len(results[0].boxes) > 0   # True = show, False = hide
+```
+
+## What This Model Does NOT Handle
+
+| Module | Status | Reason |
+|--------|--------|--------|
+| Bengali STT | ✅ Handled by `Tmanna/whisper-bengali-final` | Pre-existing model |
+| Translation | ❌ Not needed | English→Bengali not required |
+| Speaker Diarization | ⚠️ Optional | Not in core pipeline |
+| Face Tracking (DeepSORT) | ⚠️ Optional | Not in core pipeline |
+| Active Speaker Detection | 🔴 Hard/Optional | Not in core pipeline |

best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bb87b25bdc77e7641e7b768968b855cc9dce18cf03a32d0c0f565599edb5292
+size 6236010

dataset_config.yaml

@@ -0,0 +1,6 @@
+names:
+- face
+nc: 1
+path: /kaggle/working/wider_split
+train: train/images
+val: val/images

last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae322d7d7be254e99dbf0b52aa8f21cc299477503ade4d03ea3c377aded9027c
+size 6236010

metrics.json

@@ -0,0 +1,11 @@
+{
+  "mAP50": 0.6994216701471736,
+  "mAP50-95": 0.36653675270342345,
+  "Precision": 0.8555835854819701,
+  "Recall": 0.6104471858134156,
+  "dataset": "lylmsc/wider-face-for-yolo-training",
+  "base_model": "yolov8n.pt",
+  "epochs": 50,
+  "image_size": 640,
+  "hardware": "2x NVIDIA T4 (Kaggle)"
+}