Hi all. I am building a traffic counter and it requires detecting vehicles and people (drawing bounding box around them). So far I am using Yolov3 and getting above 90% accuracy for less congested areas and close to 90% accuracy even in highly congested urban areas. I was wondering if there is even a more accurate publicly available neural network? I know about Faster RCNN but so far, from what I have read, the most accurate neural network (as well as being fast) is Yolov3. Any suggestions will be appreciated. Thanks!
you can always do a fine-tuning of a pretrained model (like the faster-r-cnn of torchvision)
actually yolov3 is not the most accurate net work, it trades accuracy with speed, so you could apply it also to live streams of videos, it is not even the fastest one while yolo and yolov2 are faster and does not include detection in three layers, fast RCNN and SSD are still the most accurate ones.
Actually SSD offers the best accuracy-speed trade off and is still usable for real-time. In your case I would suggest to try it.
You can find an accuracy and speed comparison here.