Systems and methods for 3D object detection

1. A system for three-dimensional object detection, the system comprising:

one or more processors; and

a memory communicably coupled to the one or more processors and storing:

a neural network module including instructions that when executed by the one or more processors cause the one or more processors to input, to a neural network, a two-dimensional label associated with an object in an input image to produce, for the object, a two-dimensional Normalized-Object-Coordinate-Space (NOCS) image and a shape vector, the shape vector mapping to a continuously traversable coordinate shape space that combines a signed-distance-field (SDF) shape space with NOCS;

a decoder module including instructions that when executed by the one or more processors cause the one or more processors to decode the two-dimensional NOCS image and the shape vector to an object model in the continuously traversable coordinate shape space;

an initialization module including instructions that when executed by the one or more processors cause the one or more processors to:

back-project, in a frustum, the two-dimensional NOCS image to a Light Detection and Ranging (LIDAR) point cloud corresponding to the input image; and

identify one or more correspondences between the LIDAR point cloud and the object model to produce an estimate of an affine transformation between the LIDAR point cloud and the object model;

an optimization module including instructions that when executed by the one or more processors cause the one or more processors to refine iteratively the estimate of the affine transformation via differentiable geometric and visual alignment using a differentiable SDF renderer;

an autolabeling module including instructions that when executed by the one or more processors cause the one or more processors to extract automatically a three-dimensional label for the object in three-dimensional space based, at least in part, on the iteratively refined estimate of the affine transformation between the LIDAR point cloud and the object model; and

an object-detection module including instructions that when executed by the one or more processors cause the one or more processors to perform three-dimensional object detection of the object based, at least in part, on the extracted three-dimensional label for the object.

2. The system of claim 1 , further comprising a training module including instructions that when executed by the one or more processors cause the one or more processors, in accordance with a predetermined training curriculum for training the neural network, to:

use, as training data during a first iteration of a training loop, synthetic data that includes rendered driving scenes and computer-aided-design (CAD) models of a plurality of different types of vehicles with ground-truth annotations;

use, as training data during one or more iterations in a plurality of iterations subsequent to the first iteration of the training loop, a mixture of real data and the synthetic data; and

increase, over the plurality of iterations subsequent to the first iteration of the training loop, a difficulty level of the training data.

3. The system of claim 1 , wherein the two-dimensional label associated with the object is one of a two-dimensional bounding box and a two-dimensional instance mask.

4. The system of claim 1 , wherein the three-dimensional label for the object is a cuboid that encloses the object.

5. The system of claim 1 , wherein the neural network is a convolutional neural network that includes one or more fully connected layers at an end of the neural network.

6. The system of claim 1 , wherein the instructions in the optimization module to refine iteratively the estimate of the affine transformation via differentiable geometric and visual alignment using the differentiable SDF renderer include instructions to optimize the estimate of the affine transformation between the LIDAR point cloud and the object model with respect to pose, scale, and shape.

7. The system of claim 1 , wherein the object is one of a vehicle, a bicycle ridden by a bicyclist, a motorcycle ridden by a motorcyclist, and a pedestrian.

8. The system of claim 1 , wherein the system is integrated with a vehicle.

9. The system of claim 8 , wherein the vehicle is an autonomous vehicle.

10. A non-transitory computer-readable medium for three-dimensional object detection and storing instructions that when executed by one or more processors cause the one or more processors to:

input, to a neural network, a two-dimensional label associated with an object in an input image to produce, for the object, a two-dimensional Normalized-Object-Coordinate-Space (NOCS) image and a shape vector, the shape vector mapping to a continuously traversable coordinate shape space that combines a signed-distance-field (SDF) shape space with NOCS;

decode the two-dimensional NOCS image and the shape vector to an object model in the continuously traversable coordinate shape space;

back-project, in a frustum, the two-dimensional NOCS image to a Light Detection and Ranging (LIDAR) point cloud corresponding to the input image;

identify one or more correspondences between the LIDAR point cloud and the object model to produce an estimate of an affine transformation between the LIDAR point cloud and the object model;

refine iteratively the estimate of the affine transformation via differentiable geometric and visual alignment using a differentiable SDF renderer;

extract automatically a three-dimensional label for the object in three-dimensional space based, at least in part, on the iteratively refined estimate of the affine transformation between the LIDAR point cloud and the object model; and

perform three-dimensional object detection of the object based, at least in part, on the extracted three-dimensional label for the object.

11. The non-transitory computer-readable medium of claim 10 , further comprising instructions that when executed by the one or more processors cause the one or more processors, in accordance with a predetermined training curriculum for training the neural network, to:

use, as training data during a first iteration of a training loop, synthetic data that includes rendered driving scenes and computer-aided-design (CAD) models of a plurality of different types of vehicles with ground-truth annotations;

use, as training data during one or more iterations in a plurality of iterations subsequent to the first iteration of the training loop, a mixture of real data and the synthetic data; and

increase, over the plurality of iterations subsequent to the first iteration of the training loop, a difficulty level of the training data.

12. The non-transitory computer-readable medium of claim 10 , wherein the neural network is a convolutional neural network that includes one or more fully connected layers at an end of the neural network.

13. The non-transitory computer-readable medium of claim 10 , wherein the instructions to refine iteratively the estimate of the affine transformation via differentiable geometric and visual alignment using the differentiable SDF renderer include instructions to optimize the estimate of the affine transformation between the LIDAR point cloud and the object model with respect to pose, scale, and shape.

14. A method of three-dimensional object detection, the method comprising:

inputting, to a neural network, a two-dimensional label associated with an object in an input image to produce, for the object, a two-dimensional Normalized-Object-Coordinate-Space (NOCS) image and a shape vector, the shape vector mapping to a continuously traversable coordinate shape space that combines a signed-distance-field (SDF) shape space with NOCS;

decoding the two-dimensional NOCS image and the shape vector to an object model in the continuously traversable coordinate shape space;

back-projecting, in a frustum, the two-dimensional NOCS image to a Light Detection and Ranging (LIDAR) point cloud corresponding to the input image;

identifying one or more correspondences between the LIDAR point cloud and the object model to produce an estimate of an affine transformation between the LIDAR point cloud and the object model;

refining iteratively the estimate of the affine transformation via differentiable geometric and visual alignment using a differentiable SDF renderer;

extracting automatically a three-dimensional label for the object in three-dimensional space based, at least in part, on the iteratively refined estimate of the affine transformation between the LIDAR point cloud and the object model; and

performing three-dimensional object detection of the object based, at least in part, on the extracted three-dimensional label for the object.

15. The method of claim 14 , further comprising training the neural network in accordance with a training curriculum that includes:

using, as training data during a first iteration of a training loop, synthetic data that includes rendered driving scenes and computer-aided-design (CAD) models of a plurality of different types of vehicles with ground-truth annotations;

using, as training data during one or more iterations in a plurality of iterations subsequent to the first iteration of the training loop, a mixture of real data and the synthetic data; and

increasing, over the plurality of iterations subsequent to the first iteration of the training loop, a difficulty level of the training data.

16. The method of claim 14 , wherein the two-dimensional label associated with the object is one of a two-dimensional bounding box and a two-dimensional instance mask.

17. The method of claim 14 , wherein the three-dimensional label for the object is a cuboid that encloses the object.

18. The method of claim 14 , wherein the neural network is a convolutional neural network that includes one or more fully connected layers at an end of the neural network.

19. The method of claim 14 , wherein the refining optimizes the estimate of the affine transformation between the LIDAR point cloud and the object model with respect to pose, scale, and shape.

20. The method of claim 14 , wherein the object is one of a vehicle, a bicycle ridden by a bicyclist, a motorcycle ridden by a motorcyclist, and a pedestrian.