Adaptive Anchor-Box SSD for Slim Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional SSD-based object detection methods struggle with low accuracy and high computational cost when detecting slim, tightly packed objects, and require excessive resources, especially in mobile devices.
Innovation Solution
An optimized SSD model that dynamically adjusts anchor box aspect ratios and positions based on image features, using horizontal or vertical tiling, and employs depth-wise and point-wise convolution layers to reduce computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If standard SSD method uses fixed anchor boxes for object detection, then processing speed is improved, but detection accuracy for slim tightly packed objects deteriorates
Solution Approach 1:
The patent applies dynamics by making anchor box aspect ratios dynamic rather than fixed. The system automatically determines anchor box aspect ratios based on the aspect ratios of objects detected in the image, allowing the detection model to adapt to different object shapes (slim, flat, or regular objects) while maintaining processing speed. This resolves the contradiction by enabling the system to be both fast (using SSD architecture) and accurate (using adaptive anchor boxes) for various object types.
2Adaptability or versatility
If multiple anchor boxes are generated for each feature cell to detect different objects, then detection coverage is improved, but computational cost increases
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the number and aspect ratios of anchor boxes based on the detected object characteristics. Instead of using a fixed large number of anchor boxes for all cases, the system generates an appropriate number of anchor boxes with suitable aspect ratios matched to the actual objects in the image. This reduces unnecessary computational overhead while maintaining comprehensive detection coverage for different object types.
3Ease of manufacture
If square-shaped feature cells are used for uniform processing, then implementation simplicity is improved, but adaptability to different object aspect ratios deteriorates
Solution Approach 1:
The patent applies asymmetry by allowing feature cells to have different aspect ratios rather than forcing them to be square. The system determines the aspect ratio of each feature cell based on the aspect ratio of the corresponding image region and the objects within it. This enables the detection model to better match slim, flat, or regularly shaped objects while maintaining relatively simple implementation through automated aspect ratio determination.
Data Source
AI summary
Systems and methods for optimizing single shot detector (SSD) for object detection are disclosed herein. A system receives an image of a plurality of objects. Further, the system determines a plurality of feature layers and a plurality of feature cell sizes corresponding to the received image, based on an aspect ratio of the received image. Furthermore, the system determines aspect ratio of anchor boxes from trained model file, based on aspect ratio of anchor boxes, position and number of anchor boxes to be tiled in each feature cell of the plurality of feature layers. The size of the one or more anchor boxes corresponds to an anchor box aspect ratio. Additionally, the system assigns the one or more anchor boxes as a horizontal tile or a vertical tile in each feature cell, when the anchor box aspect ratio is less than a first pre-defined threshold value and greater than a second pre-defined threshold value, respectively. Further, the system generates one or more feature maps using an object detection model and a neural network (NN) model. The one or more feature maps comprises one or more feature map tensors. Furthermore, the system generates, for each layer of the one or more feature maps, a prediction tensor of predefined dimension from the one or more feature map tensors, using a prediction convolution layer. Additionally, the system detects and classifies the plurality of objects, based on the generated prediction tensor.