Source-linked AI summary
DC-SPP-YOLO: Dense Connection and Spatial Pyramid Pooling Based YOLO for Object Detection
Zhanchao Huang, Jianlin Wang, Xuesong Fu, Tao Yu, Yongqi Guo, Rutong Wang
TL;DR
YOLOv2 is fast but its accuracy is limited by weak backbone performance and incomplete use of multi-scale region features. The paper proposes DC-SPP-YOLO, combining dense convolutional connections with improved spatial pyramid pooling. On PASCAL VOC and UA-DETRAC, its mAP is reported as higher than YOLOv2, while one PASCAL VOC configuration achieves 78.4% mAP at 55.7 fps.
Problem
YOLOv2’s detection accuracy is restricted by low backbone performance and underutilization of multi-scale region features.
Method
DC-SPP-YOLO adds dense convolutional connections to YOLOv2’s backbone and uses improved spatial pyramid pooling to pool and concatenate multi-scale region features.
Results
78.4% mAP at 55.7 fps is reported for DC-SPP-YOLO 416 on PASCAL VOC 2007, while DC-SPP-YOLO’s mAP is higher than YOLOv2 on PASCAL VOC and UA-DETRAC.
Takeaways & Limitations
The reported results demonstrate the effectiveness of DC-SPP-YOLO for improving YOLOv2 detection accuracy while keeping real-time detection speed.
Takeaways & Limitations
Complex environments, large-scale variances, and rotational variations still constrain object-detection accuracy.
Abstract
from arXiv · showhide
Although the YOLOv2 method is extremely fast on object detection, its detection accuracy is restricted due to the low performance of its backbone network and the underutilization of multi-scale region features. Therefore, a dense connection (DC) and spatial pyramid pooling (SPP) based YOLO (DC-SPP-YOLO) method for ameliorating the object detection accuracy of YOLOv2 is proposed in this paper. Specifically, the dense connection of convolution layers is employed in the backbone network of YOLOv2 to strengthen the feature extraction and alleviate the vanishing-gradient problem. Moreover, an improved spatial pyramid pooling is introduced to pool and concatenate the multi-scale region features, so that the network can learn the object features more comprehensively. The DC-SPP-YOLO model is established and trained based on a new loss function composed of MSE (mean square error) loss and cross-entropy loss. The experimental results indicated that the mAP (mean Average Precision) of DC-SPP-YOLO is higher than that of YOLOv2 on the PASCAL VOC datasets and the UA-DETRAC datasets. The effectiveness of DC-SPP-YOLO method proposed is demonstrated.
2. Related Works
The related work identifies backbone feature extraction and multi-scale feature utilization as central challenges in object detection. Prior approaches improve accuracy or gradient flow through architectural connections and feature fusion, but often involve speed or small-object detection trade-offs.
- Backbone networks: YOLOv2 uses a VGG-like down-sampling convolutional backbone, but deeper convolutional layers suffer from vanishing gradients that reduce detection accuracy.The gradient becomes progressively smaller during backpropagation as activation derivatives and weights are multiplied across layers.
- Multi-scale detection: YOLOv2’s fine-grained strategy does not fully utilize multi-scale local-region features, leaving small objects difficult to detect accurately.This limitation is presented as a motivation for improving multi-scale feature processing.
- Backbone networks: Skip-connections, multi-path designs, and dense connections strengthen feature extraction or gradient propagation, but deeper backbones can severely reduce detection speed.DenseNet is described as alleviating vanishing gradients and reusing object features, while ResNet-based detectors improve accuracy at a speed cost.
- Backbone networks: DenseNet-based designs can improve accuracy while maintaining fast detection, although STDN remains slower than YOLOv2.The paper uses this speed–accuracy context to motivate applying dense connections within YOLOv2.
- Multi-scale detection: Multi-scale detection methods independently detect on multiple feature maps or fuse maps from different network layers to combine information across scales.Examples include SSD-style multi-scale detection, scale-dependent pooling, SPP, and FPN-based feature fusion.
3. Dense Connection and Spatial Pyramid Pooling Based YOLO
DC-SPP-YOLO modifies YOLOv2 with dense convolutional connections, improved spatial pyramid pooling, and a new training loss. These changes strengthen feature propagation, combine multi-scale region features, and improve reported accuracy while retaining fast detection.
- Architecture: DC-SPP-YOLO replaces part of YOLOv2’s convolutional backbone with dense-connected layers and adds a spatial pyramid pooling block.The model is constructed by modifying YOLOv2 with both components.
- Dense connection: Dense connections concatenate earlier feature maps with later-layer inputs, allowing direct feature and gradient propagation through the network.The design is intended to alleviate vanishing gradients and improve feature propagation.
- Dense connection: The DC block uses four dense connection units, each combining 3×3 and 1×1 convolutions to extract fine-grained features and reduce feature-map counts.The block is placed in a deeper backbone layer to capture richer semantic features.
- Component evaluation: DC-YOLO reached 77.6% accuracy on PASCAL VOC 2007, 0.8% above YOLOv2, while reducing detection speed by about 8.1 fps.Its complexity increased by 10.012 BFLOP/s compared with YOLOv2, while the paper reports that it retained fast detection speed.
- Motivation: The reported improvements address limited backbone feature extraction and incomplete fusion of multi-scale local region features.The paper identifies these limitations as motivations for dense connections and improved pooling.
- Spatial pyramid pooling: The improved SPP pools and concatenates multi-scale local region features while preserving feature maps instead of converting them into fixed-size vectors.Global and local multi-scale features are used together for object detection.
4. Object Detection Using DC-SPP-YOLO
DC-SPP-YOLO constructs an object-detection pipeline that preprocesses images, generates anchor boxes, extracts features, predicts boxes and classes, and suppresses redundant detections. Training combines regression and classification losses, while the model incorporates dense connections and spatial pyramid pooling.
- The object-detection process and algorithm flow are illustrated in Fig. 5 and Fig. 6.
- The pipeline augments training images, generates anchor boxes with k-means clustering, and uses IoU between boxes and centroids as the clustering distance.
- The loss function combines squared-error regression loss with binary cross-entropy classification loss, and model weights are updated iteratively until convergence.
- The convolutional network extracts object features and predicts bounding-box coordinates, confidence, and class probabilities for each grid and anchor box.
- Bounding boxes are selected using confidence and IoU thresholds, the highest-probability category is assigned, and NMS suppresses redundant boxes.
5. Experiments
Experiments evaluate DC-SPP-YOLO on PASCAL VOC and UA-DETRAC using mAP and fps, with comparisons against YOLOv2 and other detection methods. Across these datasets, the reported results show higher accuracy with only small or manageable speed differences.
- Experimental Setup: Experiments use PASCAL VOC datasets with 32,487 images across twenty categories and UA-DETRAC with 82,088 vehicle images across four categories.
- Experimental Setup: Detection accuracy is measured by mAP at IoU threshold 0.5, while detection speed is reported in fps.
- PASCAL VOC 2007: 78.4% mAP at 55.7 fps makes DC-SPP-YOLO 416 1.6% higher than YOLOv2 416 on PASCAL VOC 2007.
- UA-DETRAC: On UA-DETRAC, DC-SPP-YOLO 416 has 2.25% higher mAP than YOLOv2 416 at 57.5 fps and is reported robust in complex scenes.
6. Conclusions
DC-SPP-YOLO improves YOLOv2 by strengthening backbone feature extraction, pooling multi-scale region features, and adopting a new loss function. Experiments on PASCAL VOC and UA-DETRAC report higher detection accuracy than YOLOv2, while future work targets robustness to rotation and scale variation.
- DC-SPP-YOLO improves YOLOv2 detection accuracy while keeping real-time detection speed.
- Dense connections strengthen backbone feature extraction and alleviate the vanishing-gradient problem.
- Improved spatial pyramid pooling pools multi-scale region features within the same convolutional layer.
- A new loss function is adopted to accelerate model training.
- Detection accuracy is higher than YOLOv2 on PASCAL VOC and UA-DETRAC and is as good as popular methods on object detection tasks.
- Future research should address robust object detection under complex environments, large-scale variances, and rotational variations, including rotation and scale invariance.