search

showlab / DatasetDM

[NeurIPS2023] DatasetDM:Synthesizing Data with Perception Annotations Using Diffusion Models
https://weijiawu.github.io/DatasetDM_page/
290 stars 12 forks source link

Can't reproduce the results in other COCO-format instance segmentation dataset #25

Closed Z1zs closed 3 months ago

Z1zs commented 4 months ago

Thanks for your great job! I tried to train a P-Decoder for instance segmentation tasks in another COCO-format dataset and the mask results of P-Decoder seemed very poor. The dataset I used is CIS(Construction Instance Segmentation Dataset) and the total training epochs are 5000. Here is some results of P-Decoder in different stages(epochs), all of them seem to fail in annotating.

CIS P-Deocder

I just adjust the code for dataset loading, and all the training codes remains the same. Are there extra tricks you used in the training process like warm-up? Thank you in advance!

weijiawu commented 4 months ago

Sorry for the delayed response. The performance of instance segmentation is quite poor because instance segmentation is rather challenging. How much data did you use from CIS here?

Z1zs commented 4 months ago

thank u very much for the reply! We selected 20 images for each category in training_instances process and the total number of samples for training is 260(13 categories). Despite it's poor performance, would there be any possible methods to improve it? Thx again for your help.

weijiawu commented 4 months ago

I think that might be reasonable.

Perhaps you can try training with more data, as the AP metrics provided in our paper are also very low. Instance segmentation is very challenging when dealing with small datasets.