Open shubhank008 opened 5 years ago
Eval only works with a trimap associated with the input image. That is because the training of the model was also done with a trimap as an input. That trimap was generated from the ground truth, with erosion. Perhaps the same model with some modification could be trained without the Trimap. However, we designed this model to be used as a postprocessing step to refine background extraction and produce output image with transparency, with the first step of generating the trimap done either manually or with a segmentation technique. Sorry for the delay, and thanks for your interest!
for your pre-trained model, would there be any detriment (or have you tried) using an automated tri-map generation algo? https://ieeexplore.ieee.org/document/6694178
I understand this was probably a project for your personal use, but I found this approach for background removal much different from other better known ones like Segmentation, etc. and wanted to give it a try. Is eval/running it only possible if we have a Trimap and can I not just use any of my normal images to remove its background (I dont have trimaps for them).