Attention Network for On-Device Object Detection Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
On-device learning for autonomous vehicles faces challenges in performing positive/negative sampling and end-to-end backpropagation due to limited computing power and the absence of a teacher network, making it difficult to train object detection networks effectively without cloud access.
Innovation Solution
The method employs attention maps to improve object detection network performance by using an on-device learning device that combines feature extraction networks with attention networks, generating attentional feature maps for more accurate candidate box generation and employing cross-distilled loss layers for training, allowing for effective object detection without cloud access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If on-device learning is performed without cloud access, then autonomous vehicles can train locally using collected driving environment information, but it becomes difficult to perform positive/negative sampling and end-to-end backpropagation due to limited computing power and absence of teacher network
Solution Approach 1:
The patent divides the learning network into multiple independently trainable modules (feature extraction network, attention network, candidate box generation network, classification network). Each module can be trained separately using distilled loss functions, avoiding the need for complex end-to-end backpropagation while enabling local on-device training without cloud access.
Solution Approach 2:
The patent introduces attention maps as an intermediary component that guides the feature extraction network to focus on relevant regions. The attention network generates attention maps that are then used by subsequent networks, creating a modular training approach that simplifies the learning process and reduces computing power requirements.
2Reliability
If complex end-to-end backpropagation is performed for training, then optimal performance can be achieved, but embedded systems on autonomous vehicles have limited computing power that cannot support such complex training
Solution Approach 1:
The learning network is segmented into multiple independently trainable modules. Instead of performing complex end-to-end backpropagation across the entire network, each module (feature extraction, attention, candidate box generation, classification) is trained separately using simplified loss functions, significantly reducing computing power requirements while maintaining training effectiveness.
Solution Approach 2:
The patent changes the training approach by introducing distilled loss functions that operate on intermediate outputs (attention maps, candidate boxes) rather than requiring backpropagation through the entire network. This parameter change in the training methodology enables effective training on embedded systems with limited computing power.
3Measurement precision
If attention maps are used to improve object detection accuracy, then candidate box generation becomes more accurate, but the computational complexity and processing requirements increase
Solution Approach 1:
The patent segments the network into distinct functional modules, with the attention network being one separate component. This segmentation allows the attention mechanism to be trained independently and its outputs (attention maps) to be used by subsequent modules, improving candidate box accuracy while managing complexity through modular design.
Solution Approach 2:
Attention maps serve as intermediary representations that guide the candidate box generation process. By introducing this intermediate step, the system improves detection accuracy without requiring the final network to directly process all raw features, thereby managing computational complexity through staged processing.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for training an object detection network by using attention maps is provided. The method includes steps of: (a) an on-device learning device inputting the training images into a feature extraction network, inputting outputs of the feature extraction network into a attention network and a concatenation layer, and inputting outputs of the attention network into the concatenation layer; (b) the on-device learning device inputting outputs of the concatenation layer into an RPN and an ROI pooling layer, inputting outputs of the RPN into a binary convertor and the ROI pooling layer, and inputting outputs of the ROI pooling layer into a detection network and thus to output object detection data; and (c) the on-device learning device train at least one of the feature extraction network, the detection network, the RPN and the attention network through backpropagations using an object detection losses, an RPN losses, and a cross-entropy losses.