Abstract
Currently, object detectors based on CNN, such as RetinaNet, Faster-RCNN, and the CornerNet series, achieve high performance but suffer from significant computational costs, high model complexity, and slow detection speeds. This paper proposes Light-Net, a new lightweight object detector that utilizes a density-based approach to merge anchor boxes. To optimize speed and cost, Light-Net employs multi-scale outputs to predict objects of varying sizes. Furthermore, a novel lightweight network model is introduced, featuring "separation convolutions" to improve efficiency without sacrificing performance. Experimental results on the MS-COCO 2014 dataset demonstrate a 20.9% mAP with a model size under 10MB, showcasing the superiority of the proposed method for resource-constrained environments.
Introduction
While deep learning has revolutionized object detection, state-of-the-art models like Faster R-CNN or RetinaNet are often too computationally expensive for real-time use on edge devices or mobile platforms. The authors identify two specific bottlenecks in existing lightweight detectors: inefficient anchor generation and high model complexity. Standard convolutional layers carry a heavy parameter load, and while depthwise separable convolutions help, there is still significant room to optimize feature extraction and fusion across different scales.
Methodology
Light-Net is a one-stage detector built on three primary pillars:
Density-Based Anchor Generation
Instead of traditional K-means clustering, which is often sensitive to initial conditions, the authors introduce a density-based merging approach. By using a distance function based on $1 - IOU$ (Intersection over Union), the algorithm finds the most representative bounding box shapes. This method is more stable and provides a deterministic set of priors compared to stochastic clustering.
Separation Convolution (Sep Block)
To reduce parameters, Light-Net replaces standard $N \times N$ convolutions with a "separation" strategy, decomposing them into $N \times 1$ and $1 \times N$ layers. This reduces the calculation cost significantly while maintaining the receptive field and descriptive power of the network.
Multi-Scale Output and Focal Loss
The network predicts objects at two different scales, using high-resolution features for small objects and semantically rich, low-resolution features for larger ones. To handle the extreme imbalance between foreground and background typical in one-stage detectors, the model employs Focal Loss to prevent the detector from being overwhelmed by "easy" negative samples.
Experimental Results
The model was trained from scratch on the MS-COCO 2014 dataset. Light-Net achieved a competitive 20.9% mAP. Notably, the Light-Net32 variant has a model size of only 7.3MB (approx. 7.66 GMacs) and an inference time of roughly 25ms on an NVIDIA GTX 1070 (~40 FPS).
Qualitative Analysis
The following figures demonstrate the detection capabilities of Light-Net on complex scenes from the MS-COCO dataset. While lightweight, the model successfully identifies a variety of objects in both indoor and outdoor settings.
Conclusion
Light-Net offers a compelling trade-off between accuracy and efficiency. By rethinking anchor generation and convolutional architecture, the authors have created a detector suitable for real-time applications on mobile and edge devices. Future work will focus on increasing training iterations, applying advanced data augmentation, and exploring deeper versions of the Light-Net backbone for applications where the 10MB limit is less strict.
Summary updated