Zephyr69

Differences in model scaling could be part of the reason. Like, v6s has significantly higher flops than v5s.

How are you trying to reproduce their results? AFAIK they did not reveal their env, seeds, and exact hardware setup for training. Missing only the seed is enough to prevent...

They didn't implement their own metrics calculation methods like yolov5 but used pycocotools. You may check pycocotools repo to see what modifications are needed.

Unless either of them changed it in the most recent update, their dataset structure should be exactly the same ((images,labels) --> (train/test/valid)) or at least compatible as I've been training...

Does it train normally with yolov5? With the latest repo the mAP results on my small target dataset are still not on par with a v5 model of the same...

I don't think that exists; it has to be written specifically for yolov5. The cleanest presentation I've seen so far is to save a model, convert to .onnx, and view...

I have tested various versions of transformer/attention modules for yolov5. Some did raise a bit of metrics performance in certain types of dataset, but none of them justified the inevitable...

The functionality of the latest yolov5 is still fine with python 3.6.9 on Jetpack 4.6.x and the warning about python 3.7 can be ignored. However, it seems that nvidia-tensorrt is...

Why would you need temporal information for this tho? I thought it was pretty clear whether a person is holding an object from just one frame? Is this for cases...

There are other factors which could lead to this, despite overfitting being the most common one. You need to look at val loss to confirm the overfitting, which yolov6 doesn't...