Custom Object Detection Using YOLOv8
Budget: $12 – $30 SGD
This task focuses on Object Detection and involves training a custom object detector for unique object classes using YOLO8.
Task Requirements:
The task is to fine-tune a pretrained Object Detection model (YOLOv8) to detect 3 custom object classes. Note: The object classes must not belong to the COCO dataset.
Expected Steps to Complete the Assignment:
Data Collection and Annotation:
Images should be collected and annotated using tools such as Roboflow.
The dataset must be well-prepared for training.
Model Training and Evaluation:
The YOLOv8 model must be trained using the annotated dataset.
Model performance should be evaluated based on mean average precision (mAP).
Experimentation and Fine-tuning:
All experiments are to be logged using tools like Weights & Biases (wandb).
Model performance should be improved by fine-tuning the dataset, adjusting hyperparameters, applying data augmentation, or employing other optimization techniques.
Results from different training iterations should be compared.
Model Export and Deployment:
The trained model must be exported to a high-performance format suitable for the chosen platform (CPU or GPU) for fast inference.
Inference should be demonstrated on:
One test image
One test video (~1 minute duration)
The best-performing model should be deployed to HuggingFace Space.
Expected Submission:
A detailed discussion of the data collection and annotation process, including a link to the dataset/annotations.
An explanation of the fine-tuning process with a comparison of model performances.
Screenshots of experimental logs on Weights & Biases.
A link to the deployed model on HuggingFace Space.
Outputs from the best model:
One test image with bounding boxes.
One test video (.mp4) with bounding boxes.
This approach ensures clear documentation and demonstration of the model's performance and deployment capabilities.
Task Requirements:
The task is to fine-tune a pretrained Object Detection model (YOLOv8) to detect 3 custom object classes. Note: The object classes must not belong to the COCO dataset.
Expected Steps to Complete the Assignment:
Data Collection and Annotation:
Images should be collected and annotated using tools such as Roboflow.
The dataset must be well-prepared for training.
Model Training and Evaluation:
The YOLOv8 model must be trained using the annotated dataset.
Model performance should be evaluated based on mean average precision (mAP).
Experimentation and Fine-tuning:
All experiments are to be logged using tools like Weights & Biases (wandb).
Model performance should be improved by fine-tuning the dataset, adjusting hyperparameters, applying data augmentation, or employing other optimization techniques.
Results from different training iterations should be compared.
Model Export and Deployment:
The trained model must be exported to a high-performance format suitable for the chosen platform (CPU or GPU) for fast inference.
Inference should be demonstrated on:
One test image
One test video (~1 minute duration)
The best-performing model should be deployed to HuggingFace Space.
Expected Submission:
A detailed discussion of the data collection and annotation process, including a link to the dataset/annotations.
An explanation of the fine-tuning process with a comparison of model performances.
Screenshots of experimental logs on Weights & Biases.
A link to the deployed model on HuggingFace Space.
Outputs from the best model:
One test image with bounding boxes.
One test video (.mp4) with bounding boxes.
This approach ensures clear documentation and demonstration of the model's performance and deployment capabilities.