Warp image using optical flow

Hi, I’m trying to warp an image using a flow map(calculated using FlowNet2). I have read other similar posts in the forum which suggested using torch.nn.functional.grid_sample. I’ve never used this function before so I tried doing one based on how PWCNet used it https://github.com/NVlabs/PWC-Net/blob/master/PyTorch/models/PWCNet.py#L139

I modified it a bit to fit my data dimension but they’re basically the same implementation.


def warp(x,flo):
     B,H,W,C = x.size()
     # mesh grid
     xx = torch.arange(0,W).view(1,-1).repeat(H,1)
     yy = torch.arange(0,H).view(-1,1).repeat(1,W)
     
     xx = xx.view(1,H,W,1).repeat(B,1,1,1)
     yy = yy.view(1,H,W,1).repeat(B,1,1,1)
     
     
     grid = torch.cat((xx,yy),3).float()
    
     if x.is_cuda:
         grid = grid.cuda()

     vgrid = Variable(grid) + flo
    
     ## scale grid to [-1,1]
     vgrid[:,:,:,0] = 2.0*vgrid[:,:,:,0].clone()/max(W-1,1)-1.0
     vgrid[:,:,:,1] = 2.0*vgrid[:,:,:,1].clone()/max(H-1,1)-1.0
     
     x = x.permute(0,3,1,2)

     output = torch.nn.functional.grid_sample(x,vgrid)
     mask = torch.autograd.Variable(torch.ones(x.size()))
     mask = torch.nn.functional.grid_sample(mask,vgrid)
    
     mask[mask<0.9999]=0
     mask[mask>0]=1
    
     return output*mask

I ran the FlowNet2 to get the optical flow from the first image to the second image and then warped them using the function above. (Apparently, new users can only post one image so I can’t post the picture I’m using but it’s from the Berkeley Deep Drive Dataset)

def show(img):
    npimg = img.numpy()
    plt.imshow(np.transpose(npimg,(1,2,0)),interpolation='nearest')
    plt.show()

img0 = cv2.imread("val/picture0013.png").astype(np.float32)
img0 = np.array([img0])

flow = read_flow("work/inference/run.epoch-0-flow-field/000000.flo")
flow = np.array([flow])

img0 = torch.from_numpy(img0)
flow = torch.from_numpy(flow)

result2 = warp(img0,flow)
show(result2[0])

This is what I got

Screenshot%20from%202019-09-12%2017-14-18

Am I misunderstanding what grid_sample does??

Hi, @brinapingu,

The reason you get a poor warped image is that the function grid_sample() needs an input tensor in the range [0,1] rather than [0,255] (which is the default range of cv2). So you just have to normalize your img0. BTW, you have to be careful when warping an RGB image (cv2 reads as BRG order), img=img[:,:,::-1] may help assume the img is a numpy array with shape [H,W,3].

But I had the another question. Can anyone explain the functionality of the mask in the warp() function? In my understanding, the result of grid_sample is exactly the warped result.

Thanks in advance.

Hi, i’m facing a problem when i try to warp img0 to img1, there are many black areas around the img, I’m wondering what these area mean. And another problem is, my warped image has different color area with the ground-truth img. Any idea? thanks in advance!