[Paper Review] Scale Optimization for Full-Image-CNN Vehicle Detection

Yang Gao, Shouyan Guo+6 more|arXiv (Cornell University)|Feb 20, 2018

Advanced Neural Network Applications19 references3 citations

TL;DR

This paper proposes scale-optimized modifications to Faster R-CNN for full-image-CNN vehicle detection on the KITTI benchmark, enhancing performance through tailored anchor box selection and multi-level feature fusion. By aligning region proposal scales with the natural distribution of vehicle sizes and integrating shallow-layer features, the method improves mean average precision (mAP) from 76.3% to 83.6% on the KITTI car detection task.

ABSTRACT

Many state-of-the-art general object detection methods make use of shared full-image convolutional features (as in Faster R-CNN). This achieves a reasonable test-phase computation time while enjoys the discriminative power provided by large Convolutional Neural Network (CNN) models. Such designs excel on benchmarks which contain natural images but which have very unnatural distributions, i.e. they have an unnaturally high-frequency of the target classes and a bias towards a "friendly" or "dominant" object scale. In this paper we present further study of the use and adaptation of the Faster R-CNN object detection method for datasets presenting natural scale distribution and unbiased real-world object frequency. In particular, we show that better alignment of the detector scale sensitivity to the extant distribution improves vehicle detection performance. We do this by modifying both the selection of Region Proposals, and through using more scale-appropriate full-image convolution features within the CNN model. By selecting better scales in the region proposal input and by combining feature maps through careful design of the convolutional neural network, we improve performance on smaller objects. We significantly increase detection AP for the KITTI dataset car class from 76.3% on our baseline Faster R-CNN detector to 83.6% in our improved detector.

Motivation & Objective

• Address the performance gap in general object detectors like Faster R-CNN when applied to real-world datasets with natural scale distributions, such as KITTI.
• Improve detection accuracy for small and distant vehicles, which are underrepresented in standard benchmarks but common in real driving scenarios.
• Optimize the detector’s scale sensitivity by re-designing region proposal generation and feature extraction to match the actual scale distribution in KITTI.
• Demonstrate that adapting CNN-based detectors to domain-natural scale distributions significantly improves performance on real-world vehicle detection tasks.

Proposed method

• Revised anchor box scales in the region proposal network (RPN) to better match the natural scale distribution of vehicles in the KITTI dataset, particularly favoring smaller object scales.
• Integrated feature maps from multiple convolutional layers (e.g., 4th and 5th layers) to enhance multi-scale feature representation, especially for small vehicles.
• Employed a residual block architecture to improve gradient flow and feature diversity, enabling better learning from low-level, high-resolution features.
• Combined multi-scale proposals (using 1×1, 3×3, and 5×5 convolutional kernels) to enrich feature representation across different scales.
• Designed a hybrid network architecture that fuses shallow and deep features to improve detection sensitivity across all object scales.
• Trained the model end-to-end with careful initialization and optimization to ensure convergence and robustness on the KITTI benchmark.

Experimental results

Research questions

• RQ1 How does the performance of standard Faster R-CNN vary across different object scales in the KITTI dataset, particularly for small and distant vehicles?
• RQ2 To what extent can improving anchor box scale selection enhance detection accuracy for small-scale vehicles in real-world driving scenes?
• RQ3 Can combining feature maps from multiple convolutional layers improve detection performance on small objects without increasing computational cost?
• RQ4 How does the inclusion of residual connections and multi-scale convolutional kernels affect feature representation and detection mAP on the KITTI benchmark?

Key findings

• The proposed method significantly improves mAP on the KITTI car detection benchmark from 76.3% (baseline Faster R-CNN) to 83.6%.
• The majority of performance gains are achieved for small-scale vehicles, with detection accuracy improving substantially in the low-scale range.
• Multi-scale proposals using 1×1, 3×3, and 5×5 kernels improved initial detection performance from 64.0% to 66.3% after 10,000 iterations, though final convergence was only slightly better than baseline.
• The combination of optimized anchor boxes and multi-level feature fusion yielded the highest mAP, demonstrating the effectiveness of joint scale and feature engineering.
• Residual block integration improved performance by 0.6% and accelerated convergence, indicating better gradient flow and feature diversity.
• The model’s precision-recall curve shows marked improvement in the small-object regime, confirming enhanced sensitivity to low-scale vehicles.