PaddleDetection v2.6 Release: PP-YOLOE Family Expansion and Advanced Detection Algorithms

PaddleDetection v2.6 has been officially released after 4 months of development, introducing significant updates to Baidu's open-source computer vision detection toolkit.

Key Updates:

1. PP-YOLOE Family Expansion: New models added including PP-YOLOE-R (rotating object detection) achieving 78.14-80.73 mAP on DOTA 1.0 dataset, PP-YOLOE-SOD (small object detection) reaching 38.5 mAP on VisDrone-DET, and PP-YOLOE-DOD (dense object detection) achieving 60.3 mAP on SKU dataset. The ultra-lightweight PP-YOLOE+_t model is also introduced for edge GPU scenarios like Jetson.

2. PP-Human and PP-Vehicle: Enhanced pedestrian and vehicle analysis toolboxes with new features including reverse driving detection and lane crossing analysis. Real-time detection at 80FPS on Jetson AGX Xavier with multi-video stream support (4 channels at 20FPS).

3. Advanced Learning Methods: Open-sourced semi-supervised detection (DenseTeacher), few-shot learning (Label Co-tuning, Contrastive Tuning), and model distillation algorithms (FGD, CWD, LD) to address few-sample, generalization, annotation volume, and cold start challenges.

4. State-of-the-Art Algorithms: Added YOLOv8, YOLOv6-3.0, DINO, YOLOF, ViTDet series, CenterTrack (multi-object tracking), FCOSR (rotating detection), QueryInst (instance segmentation), and Metro3D (3D keypoint detection).

5. Infrastructure Improvements: Python 3.10 support, EMA with parameter filtering, AdamW adaptation for PaddlePaddle 2.4.1, and detection heatmap visualization.