docs.md

ImaginFusion Documentation

ImaginFusion is an application to generate 3D models based on natural language prompts. It is based on DreamFusion, but instead of using Imagen and Mip-NeRF to generate 2D priors and 3D synthesis respectively, uses Stable Diffusion and torch-ngp. It includes two interfaces, a CLI & a GUI, for better accessibility. The installation and usage instructions are in the readme file.

Table of Content

• Key External Libraries
• Key Internal Modules/Functions
  • Command Line Interface/Graphical User Interface
    • Modules Used
    • User Inputs
    • Returns
  • Args
    • Properties
  • Pipeline
    • Modules Used
    • Properties
    • Methods
  • Trainer
    • Modules Used
    • Properties
    • Methods
  • Dataset
    • Modules Used
    • Properties
    • Methods
  • Renderer
    • Modules Used
    • Properties
    • Methods
  • NeRF
    • Modules Used
    • Properties
    • Methods
  • StableDiffusionModel
    • Properties
    • Methods
  • encoder
  • TruncExp
  • softplus
  • meshDecimator
  • meshCleaner
  • getViewDirections
  • customMeshGrid
  • normalise
  • getRays
  • getCPUMem
  • getGPUMem
  • circlePoses
  • randPoses

Key External Libraries

• Tiny CUDA Neural Network Framework
• torch-ngp: A PyTorch implementation of instant-ngp
• Gradio: Hassle-Free Sharing and Testing of ML Models in the Wild
• Adan: A PyTorch implementation of Adaptive Nesterov Momentum Algorithm for Faster Optimizing by Deep Models

Key Internal Modules/Functions

Command Line Interface/Graphical User Interface

These are the Entry Point for ImaginFusion application.

Modules Used:

• Args
• Pipeline

User Inputs:

• posPrompt (str): Positive prompt for ImaginFusion.
• workspace (str): Workspace name for saving results.
• sdVersion (str): Stable Diffusion version
• hfModelKey (str): HuggingFace model key for Stable Diffusion
• fp16 (bool): Use mixed precision for training.
• seed (int): Seed value for reproducibility.
• iters (int): Number of iterations.
• lr (float): Learning Rate.
• lambdaEntropy (float): Loss scale for alpha entropy.
• maxSteps (int): Maximum number of steps sampled per ray.
• h (int): Render height for NeRF training.
• w (int): Render width for NeRF training.
• datasetSizeTrain (int): Size of training dataset.
• datasetSizeValid (int): Size of validation dataset.
• datasetSizeTest (int): Size of test dataset.

Returns:

• 3D Model (mesh.obj, mesh.mtl files and albedo.png)
• 360° Video

Args

Stores & manages all the parameters & hyperparameters for the application. All the user inputs are initially passed here before the entire Args object is passed to the Pipeline.

Properties:

• posPrompt (str): A positive text prompt.
• negPrompt (str): A negative text prompt.
• expName (str): Experiment name.
• workspace (str): Workspace directory.
• fp16 (bool): Whether to use FP16 precision.
• seed (int): Random seed for reproducibility.
• sdVersion (str): Stable Diffusion version.
• hfModelKey: High-frequency model key.
• evalInterval (int): Number of training iterations between evaluations on the validation set.
• testInterval (int): Number of training iterations between testing on the test set.
• guidanceScale (int): Guidance scale for stable diffusion.
• saveMesh (bool): Whether to save the mesh.
• mcubesResolution (int): Resolution for extracting the mesh.
• decimateTarget (float): Target for mesh decimation.
• iters (int): Number of training iterations.
• lr (float): Maximum learning rate.
• maxSteps (int): Maximum number of steps sampled per ray.
• updateExtraInterval (int): Iteration interval to update extra status.
• latentIterRatio (float): Ratio of latent iterations.
• albedoIterRatio (float): Ratio of albedo iterations.
• minAmbientRatio (float): Minimum ambient ratio.
• texturelessRatio (float): Textureless ratio.
• jitterPose (bool): Adding jitter to randomly sampled camera poses.
• jitterCentre (float): Amount of jitter to add to sampled camera pose's center.
• jitterTarget (float): Amount of jitter to add to sampled camera pose's target.
• jitterUp (float): Amount of jitter to add to sampled camera pose's up-axis.
• uniformSphereRate (float): Probability of sampling camera location uniformly.
• gradClip (float): Clip grad for all gradients.
• gradClipRGB (float): Clip grad of RGB space grad.
• bgRadius (float): Radius of the background sphere.
• densityActivation (str): Density activation function ("exp" or "softplus").
• densityThresh (float): Threshold for density grid to be occupied.
• blobDensity (float): Max density for density blob.
• blobRadius (float): Controlling the radius for density blob.
• optim (str): Optimization function.
• w (int): Render width for training NeRF.
• h (int): Render height for training NeRF.
• knownViewScale (float): Multiply h/w by this for known view rendering.
• batchSize (int): Number of images to be rendered per batch.
• bound (int): Assume the scene is bounded in box(-bound, bound)x.
• dtGamma (float): dt_gamma (>=0) for adaptive ray marching. Set to 0 to disable, >0 to accelerate rendering (but usually with worse quality).
• minNear (float): Minimum near distance for the camera.
• radiusRange (list): Training camera radius range.
• thetaRange (list): Training camera along the polar axis (up-down).
• phiRange (list): Training camera along the azimuth axis (left-right).
• fovyRange (list): Training camera fovy range.
• defaultRadius (float): Radius for the default view.
• defaultPolar (float): Polar for the default view.
• defaultAzimuth (float): Azimuth for the default view.
• defaultFovy (float): Fovy for the default view.
• progressiveView (bool): Progressively expand view sampling range from default to full.
• progressiveViewInitRatio (float): Initial ratio of the final range.
• progressiveLevel (bool): Progressively increase grid encoder's max level.
• angleOverhead (float): Overhead angle.
• angleFront (float): Front angle.
• tRange (list): Range for t values.
• dontOverrideTRange (bool): Whether to override t range.
• lambdaEntropy (float): Loss scale for alpha entropy.
• lambdaOpacity (float): Loss scale for alpha value.
• lambdaOrient (float): Loss scale for orientation.
• lambdaGuidance (float): Loss scale for guidance.
• lambdaNormal (float): Loss scale for normal map.
• lambda2dNormalSmooth (float): Loss scale for 2D normal image smoothness.
• lambda3dNormalSmooth (float): Loss scale for 3D normal image smoothness.
• H (int): Mesh height for validation.
• W (int): Mesh width for validation.
• datasetSizeTrain (int): Length of the training dataset.
• datasetSizeValid (int): Number of frames to render in the turntable video during validation.
• datasetSizeTest (int): Number of frames to render in the turntable video during test time.
• expStartIter (int): Start iteration for experiment.
• expEndIter (int): End iteration for experiment.
• writeVideo (bool): Whether to write video during testing.
• emaDecay (float): Exponential moving average decay for training NeRF.
• schedulerUpdateEveryStep (bool): Update scheduler every training step.
• refRadii (list): Reference radii.
• refPolars (list): Reference polar angles.
• refAzimuths (list): Reference azimuth angles.

Pipeline

The class for managing the entire training pipeline. In encompasses various stages of the training process, including data loading, model initialisation, and training.

Modules Used:

• NeRF
• Dataset
• Trainer
• seeder
• StableDiffusionModel

Properties:

• args (object): A configuration object containing various parameters for the pipeline.
• device: The computing device (CPU or GPU) used for training.

Methods:

• loadData: Load the dataset for training, validation and testing.
  • Inputs:
    • type (str, optional): The type of dataset to load ("train", "val", or "test"). Defaults to "train".
  • Returns:
    • DataLoader: A PyTorch DataLoader object containing the loaded dataset.
• InitiateNeRF: Initialises the NeRF model.
  • Returns:
    • NeRF: An instance of NeRf model.
• InitiateGuidance: Initialises the guidance model.
  • Returns:
    • StableDiffusionModel: An instance of guidance model.
• trainNeRF: Train the NeRF model.
  • Inputs:
    • model (NeRF): The NeRF model to be trained.
    • guidance (StableDiffusionModel): The guidance model.
    • trainLoader (DataLoader): DataLoader for training data.
    • valLoader (DataLoader): DataLoader for validation data.
    • testLoader (DataLoader): DataLoader for testing data.
• Pipeline: Starts the training pipeline by loading data, initialising models and training.

Trainer

The class for training, evaluation & testing the Text-to-3D model.

Modules Used:

• getCPUMem
• getGPUMem

Properties:

• args (object): Arguments for training.
• model (nn.Module): The neural network model.
• guidance (object): Guidance for training.
• expName (str): Experiment name.
• criterion (nn.Module, optional): Loss function. Default is None.
• optimiser (callable, optional): Optimizer for model training. Default is None.
• lrScheduler (callable, optional): Learning rate scheduler. Default is None.
• emaDecay (float, optional): Exponential moving average decay rate. Default is None.
• metrics (list, optional): List of metrics for evaluation. Default is an empty list.
• device (str, optional): Device for training (CPU or GPU). Default is None.
• verbose (bool, optional): Whether to print verbose output. Default is True.
• fp16 (bool, optional): Whether to use mixed-precision training. Default is False.
• workspace (str, optional): Workspace directory for saving logs and checkpoints. Default is "workspace".
• bestMode (str, optional): Best mode for selecting checkpoints (min or max). Default is "min".
• useLossAsMetric (bool, optional): Whether to use loss as a metric. Default is True.
• reportMetricAtTraining (bool, optional): Whether to report metrics during training. Default is False.
• useTensorboardX (bool, optional): Whether to use TensorboardX for logging. Default is True.
• schedulerUpdateEveryStep (bool, optional): Whether to update the learning rate scheduler at every step. Default is False.

Methods:

• prepareEmbeddings: Prepares text embeddings during training.
• log: Logs messages to a file.
  • Inputs:
    • args: Variable-length argument list.
    • kwargs: Arbitrary keyword arguments.
• train_step: Performs a single training step.
  • Inputs:
    • data (dict): Training data.
  • Returns:
    • Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: Predicted RGB and depth values, and training loss.
• post_train_step: Perform post-training step actions like gradient scaling and clipping.
• eval_step: Performs a single evaluation step.
  • Inputs:
    • data (dict): Evaluation data.
  • Returns:
    • Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: Predicted RGB and depth values, and evaluation loss.
• test_step: Performs a single test step.
  • Inputs:
    • data (dict): Testing data.
    • bgColor (torch.Tensor, optional): Background colour. Defaults to None.
    • perturb (bool, optional): Whether to perturb the rendering. Defaults to False.
  • Returns:
    • Tuple[torch.Tensor, torch.Tensor, None]: Predicted RGB and depth values, and placeholder.
• saveMesh: Saves a 3D mesh representation of the test predictions.
  • Inputs:
    • path (str, optional): Path to save the mesh. Defaults to None.
• trainOneEpoch: Performs one epoch of the training.
  • Inputs:
    • loader (torch.utils.data.DataLoader): DataLoader for training data.
    • maxEpochs (int): Maximum number of epochs.
• evaluateOneEpoch: Performs one epoch of the evaluation.
  • Inputs:
    • loader (torch.utils.data.DataLoader): DataLoader for evaluation data.
    • name (str, optional): Name for the evaluation. Defaults to None.
• test: Performs prediction.
  • Inputs:
    • loader (torch.utils.data.DataLoader): DataLoader for testing data.
    • savePath (str, optional): Path to save test results. Defaults to None.
    • name (str, optional): Name for the test. Defaults to None.
    • writeVideo (bool, optional): Whether to write test results as video. Defaults to True.
• train: Performs training.
  • Inputs:
    • trainLoader (torch.utils.data.DataLoader): DataLoader for training data.
    • validLoader (torch.utils.data.DataLoader): DataLoader for validation data.
    • testLoader (torch.utils.data.DataLoader): DataLoader for testing data.
    • maxEpochs (int): Maximum number of epochs.
• evaluate: Performs evaluation.
  • Input:
    • loader (torch.utils.data.DataLoader): DataLoader for evaluation data.
    • name (str, optional): Name for the evaluation. Defaults to None.

Dataset

The class for managing the data used in training and inference, including camera poses, directions, intrinsics, and more & and creates DataLoader object.

Modules Used:

• circlePose
• randPose
• getRays

Properties:

• args (object): A configuration object.
• device (str): The device on which to perform computation.
• type (str, optional): The datase type, either "train" or "all". Defaults to "train".
• H (int, optional): Height of the image. Defaults to 256.
• W (int, optional): Width of the image. Defaults to 256.
• size (int, optional): Size of the dataset. Defaults to 100.

Methods:

• collateFn: Collate function for creating batches of data.
  • Inputs:
    • idx (list): List of indices to select data for batch.
  • Returns:
    • dict: A dictionary containing batched data.
• dataLoader: Creates a DataLoader for the dataset.
  • Inputs:
    • batchSize (int, optional): The batch size. Defaults to None. If not provided, use the default batch size from args.
  • Returns:
    • DataLoader: A DataLoader object for the dataset.
• getDefaultViewData: Get data for default view(s).
  • Returns:
    • dict: A dictionary containing data for default views.

Renderer

The class for rendering 3D scenes, conducting raymarching, exporting 3D meshes, and managing density grids.

Modules Used:

• meshDecimator
• meshCleaner
• customMeshGrid
• normalise

Properties:

• args (dict): Configuration arguments.
• bound (float): The bounding box size.
• cascade (int): Number of cascades.
• gridSize (int): Size of the 3D grid.
• densityT (float): Density threshold.
• aabb_train (torch.Tensor): Training bounding box.
• aabb_infer (torch.Tensor): Inference bounding box.
• glctx: Graphics context for rendering.
• density_grid (torch.Tensor): Density grid for raymarching.
• density_bitfield (torch.Tensor): Bitfield for density grid.
• meanDensity (float): Mean density value.
• iterDensity (int): Iteration count for density updates.

Methods:

• densityBlob: Calculate density values for given points.
  • Inputs:
    • x (torch.Tensor): Input points with shape [B, N, 3].
  • Returns:
    • torch.Tensor: Density values for input points.
• forward: Placeholder.
• density: Placeholder.
• resetExtraState: Reset additional state variables.
• exportMesh: Export 3D mesh to a file.
  • Inputs:
    • path (str): Path to save the exported mesh.
    • resolution (int, optional): Resolution for mesh generation. Defaults to None.
    • decimateT (int, optional): Decimation threshold. Defaults to -1.
    • S (int, optional): Split size for mesh grid generation. Defaults to 128.
• run(raysO, raysD, lightD, ambientRatio, shading, bgColor, perturb, tThresh, binarise, **test): Perform raymarching and rendering.
  • Inputs:
    • raysO (torch.Tensor): Ray origins with shape [B, N, 3].
    • raysD (torch.Tensor): Ray directions with shape [B, N, 3].
    • lightD (torch.Tensor, optional): Light directions. Defaults to None.
    • ambientRatio (float, optional): Ambient light ratio. Defaults to 1.0.
    • shading (str, optional): Shadind mode. Defaults to 'albedo'.
    • bgColor (float or torch.Tensor, optional): Background colour. Defaults to None.
    • perturb (bool, optional): Enable ray perturbation. Defaults to False.
    • tThresh (float, optional): Threshold of t. Defaults to 1e-4.
    • binarise (bool, optional): Binarise the output image. Defaults to False.
• updateExtraState: Update additional state variables.
  • Inputs:
    • decay (float, optional): Decay factor for updating the density grid. Defaults to 0.95.
    • S (int, optional): Split size for grid generation. Defaults to 128.
• render(raysO, raysD, **kwargs): Render a scene using raymarching.
  • Inputs:
    • raysO (torch.Tensor): Ray origins with shape [B, N, 3].
    • raysD (torch.Tensor): Ray directions with shape [B, N, 3].
    • **kwargs: Additional arguments passed to the "run" method.
  • Returns:
    • dict: Rendered image, depth & weights.

NeRF

The main instantNGP model (torch-ngp). Inherits Renderer and extends from it.

Modules Used:

• Renderer
• encoder
• normalise
• truncExp
• softplus

Properties:

• nLayers (int): Number of layers in the NeRF network.
• hiddenDim (int): Hidden dimension of the NeRF network.
• encoder (tcnn.Encoding): TinyCUDA neural network encoder.
• inDim (int): Input dimension of the encoder.
• sigmaNet (Network): NeRF network for predicting sigma and albedo.
• densityActivation (function): Activation function for density prediction.
• nLayersBG (int): Number of layers in the background network.
• hiddenDimBG (int): Hidden dimension of the background network.
• encoderBG (nn.Module): Background encoder.
• inDimBG (int): Input dimension of the background encoder.
• netBG (Network): Background network.

Methods:

• forwardC: Forward pass for NeRF color prediction.
  • Inputs:
    • x (torch.Tensor): Input coordinates.
  • Returns:
    • Tuple[torch.Tensor, torch.Tensor]: Sigma and albedo predictions.
• normal: Compute surface normals.
  • Inputs:
    • x (torch.Tensor): Input coordinates.
  • Returns:
    • torch.Tensor: Surface normals.
• forward: Forward pass for NeRF rendering.
  • Inputs:
    • x (torch.Tensor): Input coordinates.
    • d (torch.Tensor): Depth values.
    • l (torch.Tensor, optional): Light direction vectors. Defaults to None.
    • ratio (int, optional): Lambertian ratio. Defaults to 1.
    • shading (str, optional): Shading mode ("albedo", "normal", "textureless"). Defaults to 'albedo'.
  • Returns:
    • Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: Sigma, colour, and normal predictions.
• density: Predict density values.
  • Inputs:
    • x (torch.Tensor): Input coordinates.
  • Returns:
    • Dict[str, torch.Tensor]: Predicted sigma and albedo.
• background: Predict background values.
  • Args:
    • d (torch.Tensor): Depth values.
  • Returns:
    • torch.Tensor: Predicted background values.
• get_params: Get network parameters and learning rates.
  • Args:
    • lr (float): Learning rate.
  • Returns:
    • List[Dist[str, Union[nn.Parameter, float]]]: List of parameter dictionaries.

StableDiffusionModel

An implementation of Stable Diffusion to generate images based on natural language propmts and act as guiding model for NeRF.

Properties:

• device (str): The device on which the model is running.
• version (str): The version of the stable diffusion model being used.
• modelPath (str): The path to the pretrained stable-diffusion model.
• precisionT (torch.dtype): The precision type for model tensors.
• vae (nn.Module): The Variational Autoencoder component of the model.
• tokenizer: The text tokenizer used by the model.
• textEncoder: The text encoder used by the model.
• unet: The UNet component of the model.
• scheduler: The diffusion scheduler used by the model.
• numSteps (int): The total number of diffusion steps.
• minSteps (int): The minimum number of diffusion steps in the specified range.
• maxSteps (int): The maximum number of diffusion steps in the specified range.
• alphas (torch.Tensor): The alpha values used in the diffusion process.

Methods:

• getTextEmbeddings: Get text embeddings for a given prompt.
  • Inputs:
    • prompt (str): The text prompt for which embeddings are to be generated.
  • Returns:
    • torch.Tensor: Text embeddings dor the input prompt.
• produceLatents: Generate latent vectors.
  • Inputs:
    • embeddings (torch.Tensor): Text embeddings.
    • h (int, optional): Height of the generated image. Defaults to 512.
    • w (int, optional): Width of the generated image. Defaults to 512.
    • numSteps (int, optional): Number of diffusion steps. Defaults to 50.
    • guidanceScale (float, optional): Scaling factor for guidance. Defaults to 7.5.
    • latents (torch.Tensor, optional): Latent vectors. Defaults to None.
  • Returns:
    • torch.Tensor: Generated latent vectors.
• decodeLatents: Decode latent vectors into images.
  • Inputs:
    • latents (torch.Tensor): Latent vectors to be decoded.
  • Returns:
    • torch.Tensor: Decoded images.
• encodeImages: Encode images into latent vectors.

  • Inputs:

    • images (torch.Tensor): Images to be encoded.

  • Returns:

    • torch.Tensor: Encoded latent vectors.
• trainStep: Perform a training step.
  • Inputs:
    • embeddings (torch.Tensor): Text embeddings.
    • predRGB (torch.Tensor): Predicted RGB images.
    • guidanceScale (int, optional): Scaling factor for guidance. Defaults to 100.
    • asLatent (bool, optional): If True, use "predRGB" as latent vectors. Defaults to False.
    • gradScale (int, optional): Scaling factor for gradients. Defaults to 1.
  • Returns:
    • torch.Tensor: Training loss

encoder

Function to create a FreqEncoder instance.

Inputs:

• inDim (int, optional): The input dimension. Defaults to 3.
• multiRes (int, optional): The degree of multi-resolution encoding. Defaults to 6.

Returns:

-Tuple: A tuple containing the frequency encoder and its output dimension.

TruncExp

Custom autograd Function for the truncated exponential operation. This function computes the exponential of input values while clamping the output to a maximum value of 15.

Inputs:

• ctx (Context): A PyTorch context object to save intermediate values for backpropagation.
• x (Tensor): The input tensor to compute the truncated exponential for.

Returns:

• Tensor: The tensor containing the truncated exponential of the input values.

softplus

Fuction to compute biased softplus activation. The softplus function is defined as softplus(x) = ln(1 + exp(x)), and this implementation allows an optional bias to be applied before computing the softplus.

Inputs:

• x (Tensor): The input tensor to apply the softplus activation to.
• bias (float): An optional bias value to be subtracted from the input tensor before applying softplus.

Returns:

• Tensor: The tensor containing the softplus activations.

meshDecimator

Function to decimate a mesh while preserving it's shape.

Inputs:

• vertices (numpy.ndarray): The vertices of the input mesh.
• faces (numpy.ndarray): The faces of the input mesh.
• target (int): The target number of faces after decimation.
• remesh (bool, optional): Whether to remesh the mesh after decimation. Defaults to False.
• optimalPlacement (bool, optional): Whether to use optimal placement during decimatin. Defaults to True.

Returns:

• Tuple[numpy.ndarray, numpy.ndarray]: The vertices and faces of the decimated mesh.

meshCleaner

Function to clean and repair 3D mesh.

Inputs:

• vertices (numpy.ndarray): The vertices of the input mesh.
• faces (numpy.ndarray): The faces of the input mesh.
• vPct (int, optional): Percentage of close vertices of merge. Defaults to 1.
• minF (int, optional): Minimum number of faces in connected components to keep. Defaults to 8.
• minD (int, optional): Minimum diameter of connected components to keep. Defaults to 5.
• repair (bool, optional): Whether to repair non-manifold edges and vertices. Defaults to True.
• remesh (bool, optional): Whether to remesh the mesh after cleaning. Defaults to True.
• remeshSize (float, optional): Target edge length for remeshing. Defaults to 0.01.

Returns:

• Tuple[numpy.ndarray, numpy.ndarray]: The vertices and faces of the cleaned and repaired mesh.

getViewDirections

Function to calculate the view direction based on the angles, the thetas, and the phis.

Inputs:

• thetas (torch.Tensor): Tensor containing theta angles in radians.
• phis (torch.Tensor): Tensor containing phi angles in radians.
• oHead (float): Angle overhead threshold in radians.
• front (float): Angle front threshold in radians.

Returns:

• torch.Tensor: A tensor of integers representing view directions.

customMeshGrid

Function to create a mesh grid for given input tensors.

Inputs:

• args: Input tensors for which the mesh grid should be created.

Returns:

• tuple: A tuple of tensors representing the mesh grid.

normalise

Function to normalise a tensor.

Inputs:

• x (torch.Tensor): Input tensor.
• eps (float, optional): A small value to prevent division by zero. Defaults to 1e-20.

Returns:

• torch.Tensor: normalised tensor.

getRays

Function to generate rays based on camera poses and intrinsics.

Inputs:

• poses (torch.Tensor): Camera poses.
• intrinsics (tuple): Camera intrinsics (fx, fy, cx, cy)
• H (int): Image height.
• W (int): image width.
• N (int, optional): Number of rays to generate. Defaults to -1, generates all rays.
• errorMap (torch.Tensor, optional): Error map for ray sampling. Defaults to None.

Returns:

• dict: A dictionary containing ray information including origins, directions, and indices.

seeder

Function to set random seed for Python, NumPy, and PyTorch.

Inputs:

• seed (int): Random seed value.

getCPUMem

Function to get current memory usage.

Returns:

• float: Current CPU memory usage in GB.

getGPUMem

Function to get current GPU usage.

Returns:

• tuple: A tuple containing the total GPU memory and GPU memory usage for each available GPU.

circlePoses

Function to generate circular camera poses.

Inputs:

• device (str): PyTorch device.
• radius (torch.Tensor, optional): Radius of the circle. Defaults to torch.tensor([3.2]).
• theta (torch.Tensor, optional): Theta angles in degrees. Defaults to torch.tensor([60]).
• phi (torch.Tensor, optional): Phi angles in degrees. Defaults to torch.tensor([0]).
• returnDirs (bool, optional): Whether to return view directions. Defaults to False.
• angleOverhead (int, optional): Angle overhead threshold in degrees. Defaults to 30.
• angleFront (int, optional): Angle front threshold in degrees. Defaults to 60.

Returns:

• tuple: A tuple containing camera poses and view directions (if returnDirs is True).

randPoses

Function to genarate random camera poses.

Inputs:

• size (int): Number of camera poses to generate.
• device (str): PyTorch device.
• args (object): Additional arguments.
• radRange (list, optional): Range for the radius. Defaults to None.
• thetaRange (list, optional): Range for theta angles in degrees. Defaults to None.
• phiRange (list, optional): Range for phi angles in degrees. Defaults to None.
• returnDirs (bool, optional): Whether to return view directions. Defaults to False.
• angleOverhead (int, optional): Angle overhead threshold in degrees. Defaults to 30.
• angleFront (int, optional): Angle front threshold in degrees. Defaults to 60.
• uniSphRate (float, optional): Rate of uniform spherical sampling. Defaults to 0.5.

Returns:

• tuple: A tuple containing camera poses, view directions, theta angles, phi angles, and radii.