Source-linked AI summary
Feature-Fused SSD: Fast Detection for Small Objects
Guimei Cao, Xuemei Xie, Wenzhe Yang, Quan Liao, Guangming Shi, Jinjian Wu
TL;DR
Small-object detection is challenging because limited resolution and information make reliable recognition difficult, while accuracy improvements often sacrifice speed. This paper adds contextual multi-level feature fusion to SSD through concatenation and element-sum modules. On PASCAL VOC2007, both modules improve mAP over baseline SSD while retaining real-time detection speed.
Problem
Small-object detection must balance limited visual information with the need for accurate and fast inference.
Method
The paper augments SSD with multi-level feature fusion that injects contextual information through concatenation and element-sum modules.
Results
1.6 and 1.7 points: concatenation and element-sum modules improve mAP over baseline SSD on PASCAL VOC2007, with testing speeds of 43 and 40 FPS, respectively.
Takeaways & Limitations
The feature-fused SSD methods achieve faster detection than DSSD with comparable accuracy while improving detection of selected small-object categories.
Abstract
from arXiv · showhide
Small objects detection is a challenging task in computer vision due to its limited resolution and information. In order to solve this problem, the majority of existing methods sacrifice speed for improvement in accuracy. In this paper, we aim to detect small objects at a fast speed, using the best object detector Single Shot Multibox Detector (SSD) with respect to accuracy-vs-speed trade-off as base architecture. We propose a multi-level feature fusion method for introducing contextual information in SSD, in order to improve the accuracy for small objects. In detailed fusion operation, we design two feature fusion modules, concatenation module and element-sum module, different in the way of adding contextual information. Experimental results show that these two fusion modules obtain higher mAP on PASCALVOC2007 than baseline SSD by 1.6 and 1.7 points respectively, especially with 2-3 points improvement on some smallobjects categories. The testing speed of them is 43 and 40 FPS respectively, superior to the state of the art Deconvolutional single shot detector (DSSD) by 29.4 and 26.4 FPS. Code is available at https://github.com/wnzhyee/Feature-Fused-SSD. Keywords: small object detection, feature fusion, real-time, single shot multi-box detector
1. INTRODUCTION
Small-object detection is difficult because objects have limited resolution and information, while existing approaches struggle to balance accuracy with real-time speed. The paper addresses this by augmenting SSD with contextual multi-level feature fusion through two modules.
- Small objects are challenging to detect because their images contain limited resolution and information.
- SSD provides fast detection by eliminating region proposals and subsequent pixel resampling.
- The proposed approach adds contextual information to SSD through multi-level feature fusion for more accurate small-object detection.
- 1.6 and 1.7 points: concatenation and element-sum fusion improve mAP over baseline SSD while operating at 43 and 40 FPS, respectively.
2. RELATED WORK
Related work introduces contextual information through combined feature maps and uses multi-scale representations for detection. SSD differs by predicting objects across multiple layers rather than relying only on top-level features.
- Context: Contextual-information methods combine features from multiple ConvNet layers or region proposals for final object prediction.
- Multi-scale representation: SSD uses multi-scale representations to detect objects at different scales and aspect ratios from multiple network layers.
- Region-based contextual methods and SSD differ in how they predict object localization and confidence.
3. MULTI-LEVEL FEATURE FUSION MODULE IN SSD
The method fuses multi-level SSD features to provide useful local context for classification and regression. It targets shallow layers that detect small objects and implements fusion through concatenation or element-wise summation.
- Multi-level feature fusion: Multi-level SSD features are extracted and fused to exploit useful local context for final classification and regression.
- Single Shot Multibox Detector: SSD predicts smaller objects from shallower layers and larger objects from deeper layers using a pyramidal feature hierarchy.
- Layer selection: The method fuses conv4_3 and conv5_3 while avoiding deeper-layer fusion for large objects to limit speed reduction and background noise.
- Concatenation module: The concatenation module upsamples conv5_3, transforms both feature maps, normalizes them, concatenates channels, and reduces dimension with a 1×1 convolution.
- Element-sum module: The element-sum module fuses two adaptively learned feature maps pointwise using equivalent weights.
4. EXPERIMENTAL RESULTS
Experiments on PASCAL VOC evaluate feature-fused SSD against SSD and DSSD for general and small-object detection, including module behavior and running speed.
- The models were evaluated on PASCAL VOC2007 and 2012, with comparisons against SSD and DSSD.
- General objects detection: 78.8 mAP for concatenation and 78.9 mAP for element-sum exceeded original SSD by 1.6 and 1.7 points, respectively.Both results were comparable with DSSD321.
- Small objects detection: 3.6 and 2.0 mAP improvements over SSD were obtained by concatenation and element-sum on 181 mostly small-object images.
- Small objects detection: Small objects with specific backgrounds improved noticeably, and co-occurring objects provided supplementary detection evidence for one another.Examples include small cars, aircraft, birds, people in cars, and bikes beside people.
- Performance comparison of two fusion modules: Learned fusion weakened background interference for occluded pot plants, whereas element-sum handled blurry cars better when context was necessary.
- Running time: 40 FPS for concatenation and 43 FPS for element-sum remained real-time and exceeded DSSD321's 13.6 FPS, with comparable accuracy.Element-sum was 3 FPS faster because it used fewer convolution kernels and layers.
5. CONCLUSION
The paper presents feature fusion in SSD for fast and accurate small-object detection, while showing that fusion modules have different strengths and that context can also add background noise.
- Feature fusion in SSD is presented as a method for fast and accurate small-object detection.
- The two fusion operations show advantages in different detection cases.
- Because context can introduce useless background noise, controlling information transmission is identified as future work.