CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting

1Harvard University, 2Columbia University, 3UC Berkeley 4Avataar.ai, 4Toyota Research Institute
Accepted to ECCV 2024

Abstract

We propose CARFF: Conditional Auto-encoded Radiance Field for 3D Forecasting, a method for predicting future 3D scenes given past observations. Our method maps 2D ego-centric images to a distribution over plausible 3D latent scene configurations and predicts the evolution of hypothesized scenes through time. Our latents condition a global Neural Radiance Field (NeRF) to represent a 3D scene model, enabling explainable predictions and straightforward downstream planning. This approach models the world as a POMDP and considers complex scenarios of uncertainty in environmental states and dynamics. Specifically, we employ a two-stage training of Pose-Conditional-VAE and NeRF to learn 3D representations, and auto-regressively predict latent scene representations utilizing a mixture density network. We demonstrate the utility of our method in scenarios using the CARLA driving simulator, where CARFF enables efficient trajectory and contingency planning in complex multi-agent autonomous driving scenarios involving occlusions.

Two-Stage Training and Inference

CARFF's two stage training process

The PC-VAE encodes images into Gaussian latent distributions. Upper right: the pose-conditional decoder stochastically decodes the sampled latent using the given camera pose into an image. The decoded reconstruction and ground truth images are used in the MSE loss for the PC-VAE. Lower right: a NeRF is trained by conditioning on latent variables sampled from the optimized Gaussian parameters. These parameters characterize the timestamp distributions derived from PC-VAE. We use a separate MSE loss for the NeRF as well.

CARFF's auto-regressive inference pipeline

The input image is encoded using pre-trained PC-VAE to obtain a latent distribution, which is fed into our MDN model. The MDN predicts a mixture of Gaussians, which is sampled to obtain a predicted latent used to render a 3D view of the scene. To perform auto-regressive predictions, we probe the NeRF for the location of the car and feed the information back to the pre-trained encoder to predict the scene at the next timestamp.

Results

Accuracy and recall curves from predicted beliefs

From the number of samples starting at 0 to 50, the belief state coverage generated under partial observation (recall), and the proportion of correct beliefs sampled under full observation (accuracy) is plotted for predicted beliefs. There is an ideal margin between the two as shown in the curves above for the two Multi-Scene CARLA datasets used to train our model.

CARFF planning with controllers

CARFF-based controllers outperform baseline controllers by choosing the optimal action in potential collision scenarios over all 30 trials conducted.


BibTeX

        
          @misc{yang2024carff,
            title={CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting}, 
            author={Jiezhi Yang and Khushi Desai and Charles Packer and Harshil Bhatia and Nicholas Rhinehart and Rowan McAllister and Joseph Gonzalez},
            year={2024},
            eprint={2401.18075},
            archivePrefix={arXiv},
            primaryClass={cs.CV}
          }