Adding New Classes to a Trained YOLO Model Without Affecting Old Weights
Add new classes to a pre-trained YOLO model, or improve existing ones, without changing the predictions of the other classes, using the RefineDetectionTrainer.
Tutorials
Some of the stuff I've learnt and want to share.
Pruning Ultralytics YOLO Models with NVIDIA Model Optimizer
Learn how to prune and fine-tune YOLO models to boost efficiency, reduce size, and speed up inference.
Tracking with Efficient Re-Identification in Ultralytics
Add reidentification to Ultralytics without compromising latency using object-level features extracted from YOLO.
Pretrain YOLO Backbone Using Self-Supervised Learning With Lightly
Use Lightly to pretrain a YOLO backbone through self-supervised learning and then fine-tune it in Ultralytics.
Boosting Inference FPS With Tracker Interpolated Detections
Boost your inference performance by exploiting Kalman filter predictions to interpolate detections for skipped frames.
Retrieving Object-Level Features From YOLO
Extract the object-level features from YOLO for downstream tasks without extra overhead.
Balance Classes During YOLO Training Using a Weighted Dataloader
Implement class balancing in Ultralytics using a weighted dataloader and improve the performance of minority class.
A Simple Trick To Increase YOLOv8's Accuracy On Small Objects With No Overhead
Get over 10% more mAP in small object detection by exploiting YOLOv8 pose models while training.
Extending YOLOv8 COCO Model With New Classes Without Affecting Old Weights
Add additional classes to pre-trained YOLOv8 model without affecting the confidences on the existing classes.
MMDetection 3.x Hacks and Tricks
Some useful tricks and hacks I learnt while using MMDetection for my research work.