* For each object's input: pos, ang - PE; siz - either PE or MLP; cla - MLP. * Then the result from these 4 are concatenated (B x nobj x

Currently all scenes_tv (test+validation) are used in training, and validation uses test. bedroom: len(self.scenes_

Load DFAUST dataset

Total number of data points

Individually generate (and normalize) data for each of batch_size number of scenes. Outputs the same as _gen_3dfront_batch_preload.

Generate x and y translation for each scene, and apply it to both the input and label of that scene.

preprocess_floor_plan must be called beforehand, as this funciton reads from the npz files it writes. Creates a new processed_<roomtype>_augm

clean_input and per_input are of the dimensions [numscene, numobj, featperobj]

For analytics purposes. traj: np array of [num_snapshot, 14, 2] representing trajectory of objects in a scene of 14 obj. traj[i] is a scene.

Denormalize x from [-1, 1] to [minimum, maximum] x has shape (pt, d). minimum/maximum have shape (d,)

Input: xyz: pointcloud data, [B, N, 3] npoint: number of samples Return: centroids: sampled pointcloud index, [B, npo

Input: xyz: input points position data, [B, C, N] points: input points data, [B, D, N] Return: ne

Embeds x to (x, sin(2^k x), cos(2^k x), ...) Different from the paper, "x" is also in the output See https://github.com/bmil

x - B, N, D

fpbpn: [batch_size, nfpbp=250, 4]

fpoc : [batch_size, maxnfpoc, pos=2], with padded 0 beyond the num of floor plan ordered corners for each scene nfpc : [bat

Input x is normalized and of dimension [numscene,numobj=7,featperobj=7=pose_d+shape_d]

Input x is normalized and of dimension [numscene,numobj=7,featperobj=7=pose_d+shape_d]

Input x is normalized and of dimension [numscene,numobj=7,featperobj=7=pose_d+shape_d]

Input x is normalized and of dimension [numscene,numobj=7,featperobj=7=pose_d+shape_d]

x : [batch_size, maxnumobj, pos+ang+siz+cla] padding_mask: [batch_size, maxnumobj], for nn.TransformerEncoder (False: not maske

:param x: [source, target] :return: [patches_idx_source, num_incident_points_target] Returns: source patches - B

Input: x - [B, N, 3] - Point cloud with batch dim as B and num points as N Returns: TFN features - F

x - B, N, 3 - Batch of point clouds that are kdtree indexed for pooling

Half of the generated scene (clean) do not have any noise added, whereas the other half (noisy) does.

allnoisy-specific: In generating the noisy positions, chairs are given greater noise while tables are given less. This is to teach the netw

halfsplit-specific: Clean pos and ang in input has no noise. Different noise levels for tables and chairs.

allnoisy-specific: input: {clean scene + gaussian gaussian noise} label: {corresponding clean scenes})

halfsplit-specific: Clean pos and ang in input has no noise.

Only makes sense for augmented dataset. Select at least 1 from each original scene, and the (n_to_select-n_original_scene) scenes are sel

nb = y.get_shape()[0].value nv = y.get_shape()[1].value nq = y.get_shape()[2].value nr = y.get_shape()[3].value nc = y.get_s

Normalize x from [minimum, maximum] to range [-1, 1]. x has shape (pt, d). minimum/maximum have shape (d,)

v: has shape [batch_size, numobj, 2], magnitude preserved, each object rotated indepndently. Returns rot_a, which has shape [batch_size, numo

x: [nobj, d], represents one scene. All elements within this array share one noise level

x: [batch_size, maxnumobj, ang_dim=2=[cos(th), sin(th)] ]

x: [B, nobj, d]