Highway Vehicle Detection Model with modified YOLOv8
Budget: ₹1,500 – ₹12,500 INR
The goal is to customize YOLO v8 with customize it's architecture then train and tune and compare with based YOLO object-detection model that accurately locates and classifies vehicles in highway scenes drawn from four separate datasets. All images depict fast-moving traffic, varied lighting, and different camera angles, so the network must generalise well across these sources.
I will supply the annotated datasets (COCO or YOLO format can be arranged). You decide whether to start from pre-trained YOLOv5, v7, or the latest YOLOv8 weights, then handle data cleaning, augmentation, training, hyper-parameter search, and evaluation. The finished model should reach a strong mAP on an unseen validation split and maintain real-time inference speed on a single NVIDIA GPU.
Deliverables
• Complete training notebook / script (PyTorch + Ultralytics or equivalent)
• Trained weights and exported ONNX/TensorRT versions
• Evaluation report: mAP, precision-recall curves, confusion matrix per class
• Brief read-me explaining environment setup and inference command
Acceptance criteria
1. ≥ 0.85 mAP@0.5 on the held-out test set across all four datasets combined
2. Inference ≥ 25 FPS on 1080p frames (RTX 3060 benchmark)
3. No critical class misses (FN rate < 5 % for ‘vehicle’) on provided test videos
The code must be reproducible end-to-end and free of proprietary dependencies. Any questions about the datasets or required metrics—just let me know.
I will supply the annotated datasets (COCO or YOLO format can be arranged). You decide whether to start from pre-trained YOLOv5, v7, or the latest YOLOv8 weights, then handle data cleaning, augmentation, training, hyper-parameter search, and evaluation. The finished model should reach a strong mAP on an unseen validation split and maintain real-time inference speed on a single NVIDIA GPU.
Deliverables
• Complete training notebook / script (PyTorch + Ultralytics or equivalent)
• Trained weights and exported ONNX/TensorRT versions
• Evaluation report: mAP, precision-recall curves, confusion matrix per class
• Brief read-me explaining environment setup and inference command
Acceptance criteria
1. ≥ 0.85 mAP@0.5 on the held-out test set across all four datasets combined
2. Inference ≥ 25 FPS on 1080p frames (RTX 3060 benchmark)
3. No critical class misses (FN rate < 5 % for ‘vehicle’) on provided test videos
The code must be reproducible end-to-end and free of proprietary dependencies. Any questions about the datasets or required metrics—just let me know.