Synthetic dataset creation for object detection and classification with deep learning

1. A computer-implemented method for building an object detection module, comprising:

obtaining mesh representations of objects belonging to specified object classes of interest,

rendering a plurality of images by a physics-based simulator using the mesh representations of the objects, wherein each rendered image captures a simulated environment containing objects belonging to multiple of said object classes of interest placed in a bin or on a table, wherein the plurality of rendered images are generated by randomizing a set of parameters by the simulator to render a range of simulated environments, the set of parameters including environmental and sensor-based parameters,

generating a label for each rendered image, the label including a two-dimensional representation indicative of location and object classes of objects in the respective rendered image frame, wherein each rendered image and the respective label constitute a data sample of a synthetic training dataset,

training a deep learning model using the synthetic training dataset to output object classes from an input image of a real-world physical environment,

deploying the trained deep learning model for testing on a set of real-world test images to generate an inference output for each test image,

adjusting the training dataset based on a success of the generated inference outputs,

identifying a “failure” image from the set of test images for which the generated inference output does not meet a defined success criterion,

feeding the “failure” image to the simulator to render additional images by randomizing the set of parameters around an environmental or sensor-based setting that corresponds to the “failure” image and to generate a respective label for each additional rendered image, and

retraining the deep learning model using the rendered additional images and the respective labels.

2. The method according to claim 1 , wherein the set of randomized parameters comprise:

number and object classes of objects present in a scene;

position and texture of objects on the table or in the bin;

texture of background including the table or bin;

settings of simulation camera; and

lighting condition of scene.

3. The method according to claim 1 , wherein generating a label for each rendered image comprises pixel-wise labeling of the respective rendered image frame, wherein each pixel is labeled based on association with an object class of interest or background.

4. The method according to claim 1 , wherein generating a label for each rendered image comprises:

determining pixels that define a respective set of corner points, each set of corner points corresponding to a bounding box around an object in the rendered image belonging to one of the object classes of interest, and

labeling each set of corner points indicative of an object class of the respective bounded object.

5. The method according claim 1 , comprising:

identifying a “success” image from the set of test images for which the generated inference output meets a defined success criterion,

feeding the “success” image and respective label from the inference output to the training dataset to retrain the deep learning model.

6. The method according to claim 1 , comprising:

for each simulated scene, rendering a plurality of images by the simulator by varying a camera pose of a simulation camera through a series of known transformations, in addition to randomizing environmental parameters, wherein the rendered images and respective labels are used for training the deep learning model,

testing the trained deep learning model on test images generated by applying said series of known transformations of camera pose for each test scene, and

determining an optimal camera pose for capturing real-world images for object detection based on identifying a continuous variation of confidence scores of inference outputs associated with the test images as a function of the applied transformations.

7. The method according to claim 1 , wherein each rendered image comprises a RGB (Red-Green-Blue) frame.

8. The method according to claim 7 , wherein each rendered image further comprises a depth frame, wherein the randomized set of parameters include environmental and sensor-based distortions obtained from a noise model of an RGB-D sensor.

9. The method according to claim 1 , wherein the mesh representations are obtained by:

scanning a real-world object belonging to an object class of interest by a sensor,

converting each frame captured by the sensor to a point cloud,

fusing the point clouds from multiple frames to a voxel grid to derive a parametrization of an object surface, and

transforming the parametrized object surface into a 3D mesh model.

10. The method according to claim 1 , wherein the mesh representations are obtained by:

downloading a mesh representation of an object belonging to an object class of interest from a mesh database, and

using the downloaded mesh representation to represent a real-world object of interest based on evaluation of a similarity metric between the downloaded mesh representation and the real-world object of interest.

11. A non-transitory computer-readable storage medium including instructions that, when processed by a computing system, configure the computing system to perform the method according to claim 1 .

12. A computing system comprising:

a processor; and

a memory storing instructions that, when executed by the processor, cause the computing system to carry out a method according to claim 1 .