Self-Calibrated Pyramid Network for Multi-Scale 3D Pillar Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pillar-based 3D object detection techniques sacrifice accuracy when dealing with objects at multiple scales, as they do not perform well with varying distances, leading to inefficiencies in computational resources.
Innovation Solution
Incorporating a self-calibrated pyramid network (SCPyramid net) with pillar-based detection, utilizing Atrous Spatial Pyramid Pooling (ASPP) blocks for channel-wise and spatial-wise attention to enlarge the receptive field and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional pillar-based detection is used, then computational efficiency is maintained, but detection accuracy deteriorates for objects at multiple scales
Solution Approach 1:
The network is segmented into multiple ASPP modules with different dilation rates (1, 2, 4, 8) that process features at different receptive fields. Each module captures spatial patterns at different scales, and their outputs are concatenated to achieve multi-scale feature fusion, resolving the contradiction between accuracy and complexity through structured segmentation.
Solution Approach 2:
The patent introduces a new dimensional aspect by applying atrous (dilated) convolutions with varying rates across multiple parallel branches. This creates a pyramid structure in the feature space that captures information at different spatial resolutions without reducing temporal or spatial dimensions, enabling multi-scale detection while maintaining computational efficiency.
2Reliability
If pillar-based detection is used, then processing speed is maintained, but detection of elements at varying distances deteriorates
Solution Approach 1:
The network dynamically adapts to varying distances through multiple ASPP modules with different dilation rates. Each module is activated based on the scale and distance of detected objects, allowing the system to dynamically adjust its receptive field without requiring fixed processing pipelines, thus maintaining both reliability and productivity.
Solution Approach 2:
The patent changes the dilation rate parameter across different ASPP modules (1, 2, 4, 8) to capture features at different spatial scales. This parameter variation enables the network to handle objects at varying distances effectively, improving detection reliability while maintaining reasonable processing speed through efficient parameter sharing.
3Area of stationary object
If Atrous Spatial Pyramid Pooling is implemented, then receptive field is enlarged, but computational complexity increases
Solution Approach 1:
The ASPP modules are designed to be universal and multi-functional, serving both feature extraction and multi-scale detection purposes simultaneously. By making the same module structure serve multiple functions across different dilation rates, the patent reduces overall computational complexity compared to implementing separate specialized modules for each scale.
Solution Approach 2:
The patent applies the concept of porous structures through atrous (dilated) convolutions that introduce holes in the convolution kernel. This porous approach allows the network to capture larger receptive fields without proportionally increasing computational complexity, as the dilated convolutions share weights and parameters across different spatial locations.
Data Source
AI summary
Techniques and systems are provided for object detection. For instance, a process can include receiving a set of 3D features, wherein the set of 3D features are generated based on an obtained 3D point cloud; downsampling the set of 3D features; pooling the downsampled set of 3D features based on Atrous Spatial Pyramid Pooling (ASPP) to generate a pooled set of 3D features; upsampling the pooled set of 3D features to generate a upsampled pooled set of 3D features; predicting bounding boxes based on the upsampled pooled set of 3D features; and outputting the predicted bounding boxes.