There seems to be an unfair comparison between the various network architectures. The reported speed and accuracy improvements should be taken with a bit of scepticism for two reasons.
* This is the first yolo implemented in Pytorch. Pytorch is the fastest ml framework around, so some of YOLOv5's speed improvements may be attributed to the platform it was implemented on rather than actual scientific advances. Previous yolos were implemented using darknet, and EfficientDet is implemented in TensorFlow. It would be necessary to train them all on the same platform for a fair speed comparison.
* EfficientDet was trained on the 90-class COCO challenge (1), while YOLOv5 was trained on 80 classes (2).
Great points, and hoping Glenn releases a paper to complement performance. We are also planning more rigorous benchmarking nonetheless.
re: PyTorch being a confounding factor for speed - we recompiled YOLOv4 to PyTorch to achieve 50 FPS. Darknet would likely top out around 10 FPS on the same hardware.
* This is the first yolo implemented in Pytorch. Pytorch is the fastest ml framework around, so some of YOLOv5's speed improvements may be attributed to the platform it was implemented on rather than actual scientific advances. Previous yolos were implemented using darknet, and EfficientDet is implemented in TensorFlow. It would be necessary to train them all on the same platform for a fair speed comparison.
* EfficientDet was trained on the 90-class COCO challenge (1), while YOLOv5 was trained on 80 classes (2).
[1] https://github.com/ultralytics/yolov5/blob/master/data/coco....
[2] https://github.com/google/automl/blob/master/efficientdet/in...