Performance of MODNet is poor when trained on 1024x1024

[hackkhai]

Hey, I read your paper and its awesome!, I trained it on 512x512 resolution, it is really good, but when i trained the same data on higher resolution like 1024x1024 (ofcourse by using resizing by short and randomcropping), but the performace difference is very noticeable. Any sugesstions on changes that i need to make the model better on higher resolution?

[ZHKKKe]

Hi, thanks for your attention. Training with 1024 resolution images doesn't seem like a good idea, because it requires a lot of training/testing resources and leads to an insufficient receptive field in the latent space. Based on my experience:

1. It may help to get better results to train the model with the input image of size 640x640
2. If you want to handle high-resolution images, you may try using the refinement module. You can choose: a) Deep Guided Filter b) The fine-tuning sub-network proposed in Background Matting V2

[hackkhai]

Thannks for your reply, I tested your Model with a different backbone and seems like the model is performing really good in high resolution network module where as the fusion result is comparetively poor in semantic module. Is it because of the Squeeze and Excitation block being used instead of E-ASPP module. if not E-ASPP can you please share how to use ASPP module? And is the poor performance in Semantic module due to very low level decoder results?

[yogeshwari-20000609]

@hackkhai I'm getting getting same issue in training for high resolution images. Can please suggest me which backbone model you used for training of high resolution images which gives you better results ?

[hackkhai]

try larger backbones like HRNet or efficientnet

[yogeshwari-20000609]

@hackkhai Thank you so much for your reply !

[yogeshwari-20000609]

@hackkhai I'm trying to load HRNet module at backbone but in wrapper function i'm getting error in reducing layers. AttributeError: 'HighResolutionNet' object has no attribute 'features' Can you please help me out in solving this issue ? Can you share your backbone architecture and weights with me. Thank you!