Compare YOLOv3, v4, and v10

someone on the Discord server asked me about YOLO V10 versus darknet YOLO so I got everything installed trained some networks and in this video I'll review the results on the first screen you'll see four videos but I'm actually comparing three different versions version 10 version three and version 4 when training YOLO V4 my training time was 115 seconds just under 2 minutes so I wanted to see what YOLO V 10 looks like when training for the exact same length of time the YOLO V10 n results weren't great It could only detect two of the five objects so I retrained YOLO V10 for a th000 Epoch instead lastly I decided to use darknet to also train one YOLO V3 tiny to see what it would look like with the same data set let's pause this video and put some text on the screen to highlight a few important things YOLO V10 n at 104 epox is in the upper left corner it took under 2 minutes to train it fails to detect three of the five objects it takes about 5.4 milliseconds per frame to predict the entire video took 13.79% corner it took over 16 minutes to train it takes about 5.3 milliseconds per frame to predict the entire video took 14.15 seconds to process darknet YOLO V3 tiny at 3,000 iterations is in the lower left corner it took three in a half minutes to train it takes less than 1 millisecond per frame to predict the entire video took 3.36 seconds to process darknet YOLO V4 tiny at 3,000 iterations is in the lower right corner it took under 2 minutes to train it takes about 1.1 milliseconds per frame to predict the entire video took 3 and 1/2 seconds to process so let's quit from this four panel video and I'll bring up a side by side comparison of YOLO V10 n and YOLO V4 tiny after the time it takes to train the first obvious difference is the time it takes to process each frame inference with ultral litic yolo V10 is nearly five times slower than with darknet yolo V4 the next difference is the confidence levels for each object found in this this video you can see how YOLO V4 tiny returns nearly 100% confidence while YOLO V10 n returns lower values and in some cases fails to DET Tech objects or incorrectly identifies duplicates the YOLO V3 tiny and YOLO V4 tiny networks trained for this video were both sized at 224 by 160 I have another video where I explain why I chose those sizes I will link to that video from the description below I don't know what size YOLO V10 n uses by default but the original videos used both to train and to run inference all measure 640 by 480 which I'm guessing should be handled easily by any YOLO framework a link to the data set I used will be in the video description I'm confident that darknet YOLO still outperforms other YOLO Frameworks in training time inference time and when it comes to the quality of the output I'm very happy with all the optimizations that have gone into darknet YOLO since the Hank AI Fork was created a year ago especially when combined with dark help and dark mark I believe that the dark Suite of YOLO applications is still the best option for people wanting to do object detection with a completely free and open- Source YOLO framework please join the darknet YOLO Discord server if you need a assistance with dark net dark help or dark mark I will leave a link in the video description below

Channel: Stephane Charette

Views: 2,725

Keywords: darknet, yolo, yolov3, yolov4, yolov10, ultralytics, neural network, object detection, computer vision

Id: 2Mq23LFv1aM

Length: 4min 38sec (278 seconds)

Published: Thu May 30 2024