Unet and masks on ppm

I’ve been working on a multiclass semantic segmentation problem using TAO unet, and vgg16 with 16 test classes.

Currently, I am successfull with 512X512 3 channel (color) images on PNG files, and 1 channel 512X512 masks on PNG files.

My problem requires over 800 classes, which will not work on PNG masks, and inspired by the TensorRT semantic segmentation quick-start example, I’m looking into PPM files for the masks.

Is that supported? How would I go about it?


In your dataset, do you know the maximum class in one PNG mask file?

The maximum number of classes in a single png file? Not yet. need to create, but it’s very expensive and time consuming so I don’t want to start unless I know that it will work

I mean TAO Unet can train multi classes dataset. For example, Mapillary Vistas dataset.
Each mask file contains multi classes.

May I know that if you already generate the ppm mask file? Can it store 800 classes?
If yes, could you share an example?

In TAO unet, indeed, since the pixel value responds to a class, it can not support more than 256 classes.

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.

Could you save to 16bit png mask image and retry? There will be 2^16 pixel values.

More, could you share an example image of 800 classes ?

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.