Validation set division on HICO-DET about upt HOT 4 CLOSED

Hi @yaoyaosanqi,

The error was caused by the incorrect data type for the input image. In the DataFactory class, you will find a transform applied to the network inputs here. This converts a PIL image to a tensor and does image normalisation. After doing the dataset split, I don't think you applied the same transform to the new datasets.

Fred.

from upt.

It ran successfully!

But there is a small issue that num-workers must be 0 to run, can I modify what settings make num-workers greater than 0 and work properly?

In the work we've done, the resnet50-based model is very close to the resnet101-based model in the non-rare category (about 0.4 mAP), but the rare class resnet101 outperforms resnet50 by about 1mAP. What could be due to? (The same goes for UPT and QPIC)
We noticed that some work replacing Resnet 50 with resnet101 was a huge boost for non-rare classes rather than rare classes. (The same goes for CDN and SDT)

Thanks！
yaoyaosanqi.

from upt.

Hi @yaoyaosanqi,

But there is a small issue that num-workers must be 0 to run, can I modify what settings make num-workers greater than 0 and work properly?

I'm not sure what's causing problems with more than one workers. Does setting the worker number to zero affect the speed?

In the work we've done, the resnet50-based model is very close to the resnet101-based model in the non-rare category (about 0.4 mAP), but the rare class resnet101 outperforms resnet50 by about 1mAP. What could be due to?

I don't have a definitive answer for this. But intuitively, having a backbone with higher depth could help extract richer features for the rare classes. The non-rare classes, on the other hand, benefit from the large amount of data, thanks to which features can already be learned effectively. This then leaves less space for improvement.

Fred.

from upt.

We have decided to temporarily set aside this issue and continue researching it in the future if needed.

Thank you!

from upt.