I have a question regarding the use of a specific model such as Deeplab and how to create a custom dataset for it.

To give a bit of background info, I want to compare panoramically stitched images of 360 degree (consisting of 6 stitched images) vs the same 6 images without stitching. Obviously you lose information by stitching these images into a single panoramic image since it contains much less pixels but it gives you a much better user experience. My goal is to find how much accuracy I lose in comparison to the speed and user experience I gain. My problem however is the lack of data, therefore I want to use a pre-trained model.

I have created a custom dataset with labelled panoramic images and labelled the 6 images of the same situation to be able to compare them equally. This dataset contains 30 panoramically labelled images and 180 regular labelled images with the same classes as Cityscapes (https://www.cityscapes-dataset.com/). Below shows an example of one of my labelled panoramic image.

Panoramic labelled image

The problem with using a pre-trained model as Cityscapes is that it is trained on straight lines, while panoramic images are spherical images placed on a 2d plane causing lines to be round and not straight. The 6 seperate images on their own are straight and will always achieve better accuracy. Basically a pre-trained model such as deeplab or any other model is biased towards my 6 images making it unfair to compare them. Therefore I want to use my custom dataset and build it on top of an existing model such as deeplab or create my own simple model.

I don't have a lot of experience with machine learning, but it feels like Deeplab is one of the only pre-trained models that I can use for my issue. I have also thought about making my own model, but I am afraid that my lack of data makes it impossible to compare them (eventhough the accuracy in my case doesn't matter, I am looking for the difference in accuracy).

My first question is: Is Deeplab the correct model to use in this situation? Or should I go with a custom model? Or are there any other models that I can use for this situation?

Custom dataset

Besides the model selection, how can I make my dataset ready for a model such as deeplab in tensorflow? All the examples that I found on deeplab make use of the same 3 datasets: ADE20K, Cityscapes and Pascal VOC 2012. None of them show examples of custom datasets.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.