ScaledYOLOv4 VS YOLOX

WongKinYiu/ScaledYOLOv4

I'm using a branch of the author's PyTorch repo

Take a look at yolov5 or scaled yolov4. They should both handle real-time training, at low enough resolution anyway; I don't know if there is any model that can do real-time detection on 4K videos. Don't pay attention to the version numbers, I think the scaled yolov4 is sliiiightly better performance.

[CVPR'21 WAD] Challenge - Waymo Open Dataset: https://waymo.com/open/challenges/2021/real-time-2d-prediction/ YOLOR (Scaled-YOLOv4-based) has the best speed/accuracy ratio on Waymo autonomous driving challenge ((Waymo Open Dataset): Real-time 2D Detection. Thanks Chien-Yao Wang from Academia Sinica and DiDi MapVision team to push Scaled-YOLOv4 further! * DIDI MapVision: https://arxiv.org/abs/2106.08713 * YOLOR https://arxiv.org/abs/2105.04206 * YOLOR-code (Pytorch): https://github.com/WongKinYiu/yolor * Scaled-YOLOv4(CVPR21): https://openaccess.thecvf.com/content/CVPR2021/html/Wang\_Scaled-YOLOv4\_Scaling\_Cross\_Stage\_Partial\_Network\_CVPR\_2021\_paper.html * Scaled-YOLOv4-code (Pytorch): https://github.com/WongKinYiu/ScaledYOLOv4 * YOLOv4: https://arxiv.org/abs/2004.10934 * YOLOv4-code (Darknet, Pytorch, TensorFlow, TRT, OpenCV…): https://github.com/AlexeyAB/darknet#yolo-v4-in-other-frameworks

The DiDi MapVision team has shown excellent results with the YOLOR and DIDI MapVision models, both based on Scaled-YOLOv4: * DIDI MapVision: https://arxiv.org/abs/2106.08713 * YOLOR https://arxiv.org/abs/2105.04206 * YOLOR-code (Pytorch): https://github.com/WongKinYiu/yolor * Scaled-YOLOv4(CVPR21): https://openaccess.thecvf.com/content/CVPR2021/html/Wang_Scaled-YOLOv4_Scaling_Cross_Stage_Partial_Network_CVPR_2021_paper.html * Scaled-YOLOv4-code (Pytorch): https://github.com/WongKinYiu/ScaledYOLOv4 * YOLOv4: https://arxiv.org/abs/2004.10934 * YOLOv4-code (Darknet, Pytorch, TensorFlow, TRT, OpenCV...): https://github.com/AlexeyAB/darknet#yolo-v4-in-other-frameworks

* Scaled-YOLOv4-code (Pytorch): https://github.com/WongKinYiu/ScaledYOLOv4

As for the BEST model, I would suggest you use Scaled YOLOv4 since it performs the best both on the cloud and edge devices.

You're out of sync, yolo-v5 is a controversial naming. And whatever side of the debate you are in state-of-the art is Scaled-Yolo v4: https://github.com/WongKinYiu/ScaledYOLOv4 https://arxiv.org/abs/2011.08036

To complicate matters further, the code for this improvement of YOLOv4 (published by one of the original YOLOv4 authors) is actually a fork of the YOLOv5 repo: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp

So I am trying to do my best and train YOLOX for an object detection case using Google Colab.

https://github.com/Megvii-BaseDetection/YOLOX I have only tried the pretrained yolo X nano. I get corner responses even if the inference image is padded with a large margin which is unexpected

* Nice implemention of Yolo that is BSD license (not GPL) https://github.com/Megvii-BaseDetection/YOLOX

I wanted to have a few choices getting localization into image search (index and search time). I immediately thought of using a region proposal network (rpn) from mask-rcnn to create patches that can also be indexed and searched (and add the localisation). I figured it might be somewhat agnostic to classes. I did not want to use mmdetection or detectron2 due to their dependencies and just getting the rpn was not worth it. I was encouraged by the PyTorch native implementations of detection/segmentation models but ended up finding yolox the best.

Megvii-BaseDetection/YOLOX

Modern video can be broken into a series of still problems. AI vision models can make these types of classification in as fast as video. Here is a particularly there is a controversial company from China that does this very well on faces in video and they have open sourced the models: https://github.com/Megvii-BaseDetection/YOLOX

Not really a problem, see results here. Just use yolox_x. Thank you for your attention.

From what I understand the swin transformer outputs a single dimension feature vector and the yolo head takes inputs from 3 different layers from the backbone?? and I think I will need to write the backbone implementation here.