Learned Deconvolutional Upsampling for Precise Object Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional methods for deconvolution in autonomous vehicles result in imprecise upsampling due to the use of static factors, leading to lower resolution datasets and less accurate object detection and tracking.

Innovation Solution

Implementing a machine-learned deconvolutional upsampling transformation in CNNs to perform deconvolution operations, which improves the spatial resolution of data structures and enhances object detection accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a static upsampling factor is used to upscale the condensed dataset, then processing efficiency is maintained, but spatial resolution and object detection accuracy deteriorate

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidspatial resolution

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies dynamics by transitioning from a static upsampling factor to a dynamic, learned upsampling transformation. The deconvolutional layer learns optimal upsampling parameters during training, allowing the system to adaptively adjust upsampling based on input characteristics while maintaining processing efficiency through automated parameter selection.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent changes parameters by replacing the fixed static upsampling factor with learnable parameters in a deconvolutional layer. These parameters are optimized during training to maximize spatial resolution and detection accuracy, enabling the system to automatically adjust upsampling behavior based on the specific characteristics of the condensed dataset.

Inventive Principle:

Principle #35Parameter changes

2Device complexity

If a static upsampling factor is used to upscale the condensed dataset, then computational simplicity is maintained, but object detection accuracy deteriorates

Engineering Contradiction:

Improvecomputational simplicityVSAvoidobject detection accuracy

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

The patent substitutes the simple mechanical approach of static upsampling with a learned deconvolutional transformation. This replacement introduces a neural network layer that automatically learns the optimal upsampling strategy, trading some computational complexity for significantly improved object detection accuracy through adaptive parameter learning.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Measurement precision

If deconvolution operations are performed with higher resolution transformations, then object detection accuracy improves, but processing time and computational resources increase

Engineering Contradiction:

Improveobject detection accuracyVSAvoidprocessing time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by performing deconvolution operations earlier in the network architecture, before final detection layers. This allows the upscaled high-resolution features to be processed in parallel with subsequent detection operations, reducing overall processing time while maintaining improved detection accuracy throughout the pipeline.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12175764B1Learned deconvolutional upsampling decoding layer

Publication Date: 2024.12.24 ZOOX INC

AI summary

Techniques for performing deconvolution operations on data structures representing condensed sensor data are disclosed herein. Autonomous vehicle sensors can capture data in an environment that may include one or more objects. The sensor data may be processed by a convolutional neural network to generate condensed sensor data. The condensed sensor data may be processed by one or more deconvolution layers using a machine-learned upsampling transformation to generate an output data structure for improved object detection, classification, and/or other processing operations.