Plugin fails when trying YOLOv4-tiny model about fastmot HOT 7 CLOSED

@GeekAlexis Thanks a lot. Will try that.

I tried your repo in Xavier-NX. It works like charm with Good FPS and accuracy! Good job

from fastmot.

I am sorry. I found my mistake. The tiny model had only been trained on 80 classes, not 90, so my NUM_CLASSES setting was incorrect.

By the way, I am getting speeds of around 16 FPS with only 4-5 objects onscreen. I am using a Jetson Nano. Is it possible to optimize this further and get it closer to 30 FPS?

Once again, thank you for the great work.

from fastmot.

The detections using YOLOv4-tiny are quite poor as well. I am only able to detect 1 out of several people in the video, and even that was a partial bounding box, not a complete one. Is there any way to improve this?

from fastmot.

@arsenal-2004 Yes there are a few things you can do to optimize:

1. Train your YOLOv4-tiny on a pedestrian dataset. The pre-trained COCO model performs poorly on small objects.
2. Use a smaller batch size like 8 for feature extractor and/or get a faster ReID model like MobileNetv2.
3. Disable feature extractor. This requires modifying the code a little bit.

FYI, Jetson Nano is too constrained for an involved task like this. If you want it real-time on Nano, expect to see a performance drop.

from fastmot.

@GeekAlexis Thank you so much for your reply. I'll try out your suggestions and see if the performance improves.

from fastmot.

@GeekAlexis , I was also trying to implement YOLOv4-Tiny in Xavier-NX. I have a Yolov4-Tiny trained with one class. I have converted this model to onnx and TensorRT using the link you have mentioned - https://github.com/jkjung-avt/tensorrt_demos

My doubt is, Do i need to train a ReID model also on the same class to implement tracking? If yes, is there any specific steps for yolov4-tiny?

from fastmot.

@PiyalGeorge Yes, you need to train a ReID model on the same class, assuming the class is not person. And there is nothing specific for tiny because ReID is independent of detector. FYI, you shouldn't convert YOLO to TensorRT, converting to ONNX is enough.

from fastmot.