[SOLVED] Is it possible to get the number of pixels of the detected object with Detectron2?

Is it possible to get the number of pixels of the detected object with Detectron2?

I'm totally new to deep learning. I have a Detectron2 model detecting solar panels located on rooftops. I'd love to calculate the surface of the detected panels, therefore I need to get the number of pixels of detected objects. Is there any way to do it?

Solution

If you use official framework and segmentation over detections (for more precise surface estimation) on image im like:

outputs = predictor(im)

Dict outputs has the key instances (you can see source code) - you can just take the masks boolean tensors and calculate its sum:

masks = outputs['instances'].pred_masks
masks.shape

# torch.Size([15, 480, 640])

So we have here 15 objects and boolean masks size of input image as masks for each object, now we calculate total amount of pixels for each detected object:

torch.sum(torch.flatten(masks, start_dim=1),dim=1)

# tensor([32251,  8786,  2513, 11821,  1395,  2129,   496,  2722,  1020,   571,
     1060,   806,  2006,   535,   297], device='cuda:0')