Some question about train and validation and inference

[okt7760]

Hi @fcakyon ,I have some trouble in training about the detail of Training details,so I was hoping you could offer some help in achieving the results you've announced
The first question concerns the processing of visdrone datasets,the val dataset is need to be slide,slice_visdrone.py use in val or training or testing?
Second question,i am use your checkpoints from :
https://github.com/fcakyon/sahi-benchmark/releases/download/v0.0.2/tood_sf_visdrone.pth (TOOD+SF+SAHI)
and config use :
https://github.com/fcakyon/small-object-detection-benchmark/blob/main/mmdet_configs/visdrone_tood/tood_crop_480_960_cls_60.py
use the code
https://github.com/fcakyon/small-object-detection-benchmark/blob/main/eval_tools/predict_evaluate_analyse.py
we have very poor result
We find the console constantly display ignore unverified…………，so we consider the problem of visdrone, but we use your tool visdrone_to_coco.py and no changes were made
hope you can help me these two problems

[fcakyon]

The first question concerns the processing of visdrone datasets,the val dataset is need to be slide,slice_visdrone.py use in val or training or testing?

You have to only convert the visdrone training and validation splits into coco format using visdrone_to_coco. No need to slice the dataset if you use the config files from this repo.

[okt7760]

Thank you for your reply!

The Validation splits do not require slice with slice_visdrone scirpt ?

I used the slice visdrone validation split in the training process,Is this a wrong step?

For validation split operations, only coco format is required, and slice is not required?

[fcakyon]

The Validation splits do not require slice with slice_visdrone script ?

Yes, it does not require.

I used the slice visdrone validation split in the training process,Is this a wrong step?

Yes, it is wrong.

For validation split operations, only coco format is required, and slice is not required?

Yes, it is true.

[fcakyon]

Second question,i am use your checkpoints from : https://github.com/fcakyon/sahi-benchmark/releases/download/v0.0.2/tood_sf_visdrone.pth (TOOD+SF+SAHI) and config use : https://github.com/fcakyon/small-object-detection-benchmark/blob/main/mmdet_configs/visdrone_tood/tood_crop_480_960_cls_60.py use the code https://github.com/fcakyon/small-object-detection-benchmark/blob/main/eval_tools/predict_evaluate_analyse.py we have very poor result We find the console constantly display ignore unverified…………，so we consider the problem of visdrone, but we use your tool visdrone_to_coco.py and no changes were made

You are not slicing the visdrone validation split, right?

You convert the validation split with visdrone_to_coco script, right?

What is your sahi version?

[okt7760]

Thanks your reply!
sahi version : 0.11.0
First i am use the slice_visdrone script to slice validation file (VisDrone2019-DET-val) after that convert the validation with visdrone_to_coco script,Is this the right step ?

[fcakyon]

It is wrong. You have to convert the validation split without slicing it. No manual slicing is required.

[okt7760]

According to your reply, I have revised it and it is close to the result on the paper. Thanks for your help to solve my problem 😊 !!

[fcakyon]

I am glad to hear that!