Trying to re-train planercnn with custom data

[rjose97]

Hello friend,

Thanks a lot for your project.

I’m trying to do planar segmentation of single RGB image using your deep neural architecture but sometimes I don’t obtein good results. So, I’d like to re-train it, or train again, with some custom data, where I have plane parameters for each frame. However, I need the right segmentations for each plane. I can share with you my custom data in order to improve your final results.

[dLopes-SE]

Hello @rjose97,

You should change ScanNet class in order to load your files and, instead of only loading one .npy file with all the planar equations in world coordinates (like PlaneRCNN does), you need to load a .npy or a .txt, in my case, for each frame. Then you need to associate each segmentation to each plane, which is your question. You can do it with something like this:

newPlanes = []  
        newPlaneInfo = []
        newSegmentation = np.full(segmentation.shape, fill_value=-1)
        newIndex = 0

        for oriIndex, count in segmentList:
            # Non-planar label
            if oriIndex == 0:
                continue

            # Case plane does not exist (due to an error in Matlab)
            if np.amax(planes[oriIndex-1]) == 0:
                continue

            newPlanes.append(planes[oriIndex-1])   
            newSegmentation[segmentation == oriIndex] = newIndex
            #newPlaneInfo.append(self.plane_info[oriIndex] + [oriIndex])
            newIndex += 1
            continue

segmentation = newSegmentation
planes = np.array(newPlanes)
plane_info = newPlaneInfo   

Note: My data generator has an error is some unique frames, that's why I'm checking it.

I hope it was usefull, good luck.

Greetings, Dylan Lopes, UC, Portugal.

[rjose97]

Thank you so much

[Karthikaddagadderamesh]

@dLopes-SE @rjose97 Hello I plan to use network to train on my own dataset. I already have a dataset which contains RGB images,semantic labels images and RGBD images. I would like to know if I could create planes and plane normal for my own dataset and use it.

Any information is welcomed Thanks in advance