Display logs with respect to the log frequency

This function returns samples for current contexts

This function sets title for tensorboard plot.

This function computes 2d convolutions followed by the specific nonlinear function(LeakyReLU, ELU, or None)

This function plots images on tensotboard.

This function masks losses while avoiding zero division.

This function calculates photometric reconstruction loss (0.85*SSIM + 0.15*L1)

This function packs camera dimension to batch dimension.

This function prints various metrics for depth estimation accuracy.

This function unpacks batch dimension into batch and camera dimensions.

This function visualizes disparity map using colormap specified with disparity map variable.

This function computes reprojection loss using auto mask.

This function computes edge-aware smoothness loss for the disparity map.

This function constructs datasets.

This function logs depth and pose information for monitoring training process.

This function plots normalized images on tensotboard.

Pass a minibatch through the network and generate images, depth maps, and losses.

This function projects voxels into 2D image coordinate. [b, feat_dim, n_voxels] -> [b, feat_dim, d, h, w]

This function resizes images for tensorboard plot.

This function reorganize samples to match our trainer configuration.

This function back-projects 2D image points to 3D.

This function calculates depth error using various metrics.

This function computes depth metrics, to allow monitoring of training process on validation dataset.

This function computes spatial loss.

This function computes spatio-temporal loss.

This function computes 1d convolutions follwed by the nonlinear function(LeakyReLU, or None)

This function evaluates models on full validation dataset.

This function follows structure of dgp_dataset/generate_depth_map in packnet-sfm. Due to the version issue with dgp, minor revision

This function returns samples for backward and forward contexts

This function returns mean and standard deviation of the overlapped features.

obtain normalized warped images using the mean and the variance from the overlapped regions of the target frame.

This function computes pose for a single camera.

This function warps source image to target image using backprojection and reprojection process.

Log less frequently after the early phase steps

This function sets training process.

This function logs outputs for monitoring using tensorboard.

This function plots disparity maps on tensotboard.

This function transforms disparity value into depth map while multiplying the value with the focal length.

This function trains models.

This function transforms masks to match input rgb images.

This function augments camera parameters for depth synthesis.

This function augments depth estimation results using augmented extrinsics [batch, cam, 4, 4]

This function backprojects 2D features into 3D voxel coordinate using intrinsic and extrinsic of each camera. Self-occluded regions a

This function calculates pixel coords for each point([batch, n_voxels, 1, 2]) to sample the per-pixel feature.

This function creates valid mask in voxel coordinate by projecting self-occlusion mask to 3D voxel coords.

This function transforms camera name list to indices

This function computes depth augmentation loss(consistency, smoothness).

This function computes auto mask using reprojection loss and identity reprojection loss.

This function computes depth map for each viewpoint.

This function calculates edge-aware smoothness.

This function computes losses.

This function computes multi-camera posse in accordance with the network structure.

This function computes pose consistency loss in "Full surround monodepth from multiple cameras"

This function computes spatio & spatio-temporal transformation for images from different viewpoints.

This function calculates SSIM loss between predicted image and target image.

output: [batch, 3, num_depths, height * width]

output: [batch, 3, height * width]

output: [batch, 3, z_dim, y_dim, x_dim] [b, :, z, y, x] contains (x,y,z) 3D point

This function distrubutes pose to each camera by using the canonical pose and camera extrinsics. (default: reference camera 0)

This function sets dataloader for validation in training.

This function returns depth map for nuscenes dataset, result of depth map is saved in nuscenes/samples/DEPTH_MAP

This function is defined to meet dgp version(v1.4)

This function returns relative camera indices from given camera list

This function transforms pose information in accordance with DDAD dataset format

This function backward-warp source depth into the target coordinate. src -> target

This function logs outputs for visualization.

This function prints epoch, iteration, duration, loss and spent time.

This function renders projected images using camera parameters and depth information.

This function predicts disparity maps.

This function predicts poses.

This function applies 1x1 convolutions to features from non-overlapping features.

This function applies 1x1 convolutions on overlapping features. Camera configuration [0,1,2] or [0,1,2,3,4,5]:

This function prints amount of time taken in user friendly way.

This function reprojects transformed 3D points to 2D image coordinate.