Meeting date: 06-04-2023 (10h30)
Last Thursday I had problems with my computer's graphics card. Since I couldn't get PyTorch to recognize the graphics card, and not to waste any more time, I formatted the computer.
I have been looking for other multitasking models that could detect more object classes. Similar to YOLOP, I only found 3 networks - YOLOP, YOLOPv2 and HybridNets - Any of these only detect cars, buses, trucks, and trains. There is also the Detectron2, that seems more capable than YOLOP, but I didn’t have time to learn how this one works.
Almost every YOLO version have a second model, where you can extract segmentation and bounding boxes, but I'm not sure if it is considered multi-tasking, since I find schemes with a single head.
I also saw another model that seemed interesting, because it does several types of segmentation simultaneously - OneFormer I don't know if I'm using the wrong keywords, but I can't find anything super relevant. I think that I need to study more about multitask networks to be able to do a better search.
A few days ago, I finished the YOLOv5 (detections only) module. The problems that I had in the last week were easy to solve. Here is a demonstration:
Not so easy was the segmentation part, but the solution turned out to be somewhat simple:
https://www.youtube.com/watch?v=D1dNi_dDUts
In this demo, only two classes (cars and pedestrians) are illustrated for simplicity in the receiver - Previously I only had the drivable area and the lines (2 classes)
I started to look to the code of YOLOv8, but I didn't make any progress yet