Compare YOLOv4-tiny and YOLOv7-tiny

I recently started looking at YOLO V7 tiny all of my prior projects typically use YOLO V3 V4 or one of the other tiny variants training with dark net and YOLO V7 tiny is quite a bit slower than what I'm used to on the exact same data set and Hardware Nvidia RTX 3090 YOLO V4 tiny at 768 by 576 took three hours to train when Max batches was set to 20K meanwhile YOLO V7 tiny at the same resolution took nine hours to train when Max batches was set to 20K the charts look significantly different when training with the 80 class Ms Coco data set the claim is that YOLO V7 is both faster and more precise than the equivalent YOLO V4 configuration as you can see from this video with my 19 class playing card data set the results look very similar on the left YOLO V4 tiny takes approximately four milliseconds for inference on the right you'll know V7 tiny takes approximately five milliseconds for inference so in my case YOLO V4 tiny is still faster than the yellow V7 tiny I'm using RTX 3090 Cuda 11.8 and CU DNN 8.4 all running on Ubuntu 2004. I won't pause the video on all the differences but here are some obvious ones with yolo V4 tiny the confidence levels are at or close to 100 percent while the ones with yolo V7 tiny don't quite reach 100 percent on blurry objects YOLO V4 tiny is picking things up correctly even though I did not annotate any blurry objects while on YOLO V7 tiny some mistakes are made or objects are being missed I was using darknet from Alexi ab's GitHub repo to train not the Python scripts from one can you apologies if I mispronounced your name I'm still hopeful that Alexia B will modify darknet to bring the necessary changes into the C C plus plus code base I am uncertain if this may be the reason why I'm not seeing measurable improvements in any of my neural networks when I switch from V4 to V7 as always I hope this video was useful if you want to discuss further please join the Dark Net YOLO Discord a link is included in the video description below

Channel: Stephane Charette

Views: 4,420

Keywords: darknet, yolo, neural network, computer vision, object detection

Id: JSgDs0XXz8M

Length: 3min 11sec (191 seconds)

Published: Tue Nov 01 2022