Wrapper for grid_sample, uses pixel coordinates

Args: max_shape (tuple): for 1/8 featmap, the max length of 256 corresponds to 2048 pixels

Q: [BH1W1, 1, dim] K: [BH1W1, H3W3, dim] Q_r: [BH1W1, H3W3, dim] K_r: [BH1W1, H3W3, dim]

Returns: scheduler (dict):{ 'scheduler': lr_scheduler, 'interval': 'step', # or 'epoch' }

Photometric augmentation

Create submission for the Sintel leaderboard

cost_maps - B*H1*W1, cost_heads_num, H2, W2 coords - B, 2, H1, W1

Occlusion augmentation

Create the optimizer and learning rate scheduler

Applies the flow color wheel to (possibly clipped) flow components u and v. According to the C++ source code of Daniel Scharstein Accord

Estimate optical flow between pair of frames

Wrapper for grid_sample, uses pixel coordinates

Flow is represented as difference between two means flow = mean1 - mean0

Flow is represented as difference between two coordinate grids flow = coords1 - coords0

Generates a color wheel for optical flow visualization as presented in: Baker et al. "A Database and Evaluation Methodology for Optical F

Read .flo file in Middlebury format

Loss function defined over sequence of flow predictions

Upsample flow field [H/8, W/8, 2] -> [H, W, 2] using convex combination

Upsample flow field [H/8, W/8, 2] -> [H, W, 2] using convex combination

Perform evaluation on the FlyingChairs (test) split

Peform validation using the KITTI-2015 (train) split

Peform validation using the Sintel (train) split