The original author of YOLO stopped working on it[1]. Alexey Bochkovskiy, aka AlexeyAB, created a fork on GitHub and wrote an extensive guide to customizing YOLO's network architecture, added new features, and has answered zillions of questions.<p>1: <a href="https://twitter.com/pjreddie/status/1230524770350817280" rel="nofollow">https://twitter.com/pjreddie/status/1230524770350817280</a>
I object to the use of the word "optimal" for a task like object detection; it feels counterproductive to claim that this is the "optimal" way of solving such a broad and complex problem. Great results, but their language needs some tempering.
A video on YOLOv4 - Really informative.
<a href="https://www.youtube.com/watch?v=_JzOFWx1vZg" rel="nofollow">https://www.youtube.com/watch?v=_JzOFWx1vZg</a>
I have had a lot of fun working with YOLO v3 for robotics applications, very excited to try these updates. Thanks to the authors for the updates and good documentation. Good object recognition is the backbone of a huge range of future applications, and YOLO has been a good option for a while.
I'm a little skeptically of the Swish implementation after looking at Table 2.<p>Method | Top 1 | Top 5
No-op | 78% | 94%
Swish | 64.5% | 86%
Mish | 79% | 94.5%<p>Swish is the only value that decreases performance (and by a huge magnitude) but a very related methodology improves performance hummm...
<a href="https://medium.com/@riteshkanjee/yolov4-superior-faster-more-accurate-object-detection-7e8194bf1872" rel="nofollow">https://medium.com/@riteshkanjee/yolov4-superior-faster-more...</a><p>Article on YoloV4