PaperlayerView exact revision on arXiv

Source-linked AI summary

HIC-YOLOv5: Improved YOLOv5 For Small Object Detection

Shiyi Tang, Shu Zhang, Yini Fang

arXiv:2309.16393v2cs.CV

TL;DR

Small object detection in UAV imagery is difficult, while existing improvements can impose high computation costs. HIC-YOLOv5 adds a small-object prediction head, involution-based channel fusion, and CBAM attention; on VisDrone2019, it improves mAP@[.5:.95] by 6.42% and mAP@0.5 by 9.38% versus YOLOv5s.

  • Problem

    Small objects are difficult to detect in UAV imagery, and existing improvements can have high computation costs that hinder real-time deployment.

  • Method

    HIC-YOLOv5 combines a higher-resolution small-object prediction head, an involution block between the backbone and neck, and lightweight CBAM attention at the backbone output.

  • Results

    6.42% mAP@[.5:.95] and 9.38% mAP@0.5 improvements were achieved over YOLOv5s on VisDrone2019.

  • Takeaways & Limitations

    HIC-YOLOv5 improves small-object detection performance while using CBAM to reduce computation cost and emphasize important channel and spatial information.

Abstract

from arXiv · show

Small object detection has been a challenging problem in the field of object detection. There has been some works that proposes improvements for this task, such as adding several attention blocks or changing the whole structure of feature fusion networks. However, the computation cost of these models is large, which makes deploying a real-time object detection system unfeasible, while leaving room for improvement. To this end, an improved YOLOv5 model: HIC-YOLOv5 is proposed to address the aforementioned problems. Firstly, an additional prediction head specific to small objects is added to provide a higher-resolution feature map for better prediction. Secondly, an involution block is adopted between the backbone and neck to increase channel information of the feature map. Moreover, an attention mechanism named CBAM is applied at the end of the backbone, thus not only decreasing the computation cost compared with previous works but also emphasizing the important information in both channel and spatial domain. Our result shows that HIC-YOLOv5 has improved mAP@[.5:.95] by 6.42% and mAP@0.5 by 9.38% on VisDrone-2019-DET dataset.

I. INTRODUCTION

Small-object detection in UAV imagery is difficult because altitude makes targets small, causing occlusion, low density, and lighting variation. HIC-YOLOv5 addresses this with a dedicated high-resolution prediction head, involution-based channel enhancement, and CBAM attention while reducing computation cost.

  • UAV imagery contains predominantly small objects at high altitude, making detection vulnerable to occlusion, low target density, and dramatic lighting changes.
  • HIC-YOLOv5 adds a Small Object Detection Head that detects small objects using higher-resolution feature maps.Higher resolution makes tiny-object features easier to extract.
  • 6.42% mAP@[.5:.95] and 9.38% mAP@0.5 improvements are reported on the VisDrone dataset over YOLOv5.
  • An involution block bridges the backbone and neck to increase feature-map channel information.
  • CBAM is applied at the backbone end to emphasize important channel and spatial features with lower computation cost than prior work.

II. RELATED WORKS

YOLOv5 is presented as a computationally efficient one-stage detector suited to real-time use. Its architecture extracts and fuses multiscale features before prediction through three resolution-specific heads.

  • YOLOv5 is described as suitable for real-time object detection because of its performance and computational efficiency.
  • YOLOv5-6.0 uses a backbone for feature extraction, a neck for feature fusion, and prediction heads for final detection.
  • The backbone combines Conv, CSPDarkNet53, and SPPF modules to learn and enrich multiscale feature representations.
  • The neck draws on FPN and PANet to combine semantic and positional information across feature-map levels.
  • Three prediction heads process 80 × 80, 40 × 40, and 20 × 20 feature maps for large, medium, and small objects, respectively.

B. Previous Works on Small Object Detection

Prior small-object methods modify YOLOv5’s feature-fusion architecture, often increasing computation cost. HIC-YOLOv5 instead adds targeted changes: a high-resolution head, an involution block, and CBAM.

  • Previous Works on Small Object Detection: Prior work improves small-object detection by replacing or redesigning YOLOv5’s neck and feature-fusion structure.
  • Previous Works on Small Object Detection: Changing the entire neck can improve feature fusion but results in larger computation cost.
  • HIC-YOLOv5: HIC-YOLOv5 adds a high-resolution prediction head specifically for small and tiny objects.
  • HIC-YOLOv5: HIC-YOLOv5 adopts an involution block at the neck entrance and incorporates CBAM into the backbone.

A. Convolutional Block Attention Module(CBAM)

The method places CBAM in the backbone rather than the neck to highlight meaningful channel and spatial features while limiting computation on large feature maps. The section also describes involution as a channel-shared, spatially varying operation that enriches receptive-field information.

  • CBAM placement: CBAM is placed in the backbone to highlight significant features during extraction instead of generating feature pyramids in the neck.Using CBAM in the neck can increase parameters and computation because connected feature maps may be large.
  • CBAM placement: The CBAM input feature map is 20 × 20, 32 times smaller than the 640 × 640 image, reducing the associated computation cost.This size relationship is given as the reason the backbone placement does not impose a large computational burden.
  • CBAM operation: CBAM generates channel and spatial attention maps that refine the input feature map, emphasizing meaningful information and suppressing redundant information.Its channel module uses global max- and average-pooling, while its spatial module pools values at corresponding pixel positions across feature maps.
  • Involution: Involution uses kernels that vary across spatial positions but are shared across channels within groups.For each pixel, a specific K × K kernel is designed and applied across the channels.
  • Involution: Involution implicitly disperses channel information from one pixel into its spatial vicinity, enriching receptive-field information.The operation is presented as a bridge between feature representations whose channel information can otherwise be reduced by a 1 × 1 convolution.

C. Prediction Head

YOLOv5’s existing prediction-head resolutions can make tiny-object features difficult to preserve as network depth reduces feature-map resolution. The proposed small-object head uses higher-resolution feature maps, while each head ultimately predicts box geometry, objectness, and class information.

  • Existing prediction heads: YOLOv5’s three prediction heads use 80 × 80, 40 × 40, and 20 × 20 feature maps for large, medium, and small objects, respectively.The different resolutions support multiple object sizes but can make tiny-object detection difficult.
  • Motivation: Tiny-object features can be ignored because deeper convolutional blocks reduce feature-map resolution, making those few-pixel features difficult to extract.The section motivates a dedicated head for feature maps with larger resolution.
  • Small Object Detection Head: The Small Object Detection Head is added to detect small objects from higher-resolution feature maps.Higher resolution makes features of tiny and small objects easier to extract.
  • Head outputs: Each prediction head outputs bounding-box coordinates and size, objectness confidence, and the object class.Anchors form candidate boxes before the final bounding boxes are generated.

D. Loss Function

HIC-YOLOv5 uses a composite loss for objectness, bounding-box regression, and class probability, alongside data augmentation intended to improve robustness. The experiments report excellent detection accuracy on VisDrone2019.

  • Loss composition: The HIC-YOLOv5 loss combines objectness, bounding-box, and class-probability losses.The loss is expressed as a weighted sum of these three components.
  • Loss components: Binary cross entropy is used for objectness and class probability, while CIoU loss is used for bounding-box regression.These choices assign classification-style losses to objectness and class prediction and CIoU to box regression.
  • Data augmentation: YOLOv5 data augmentation includes Mosaic, Copy paste, Random affine, MixUp, HSV augmentation, and Cutout, with added center cropping.The paper describes data augmentation as an essential technique for enhancing model robustness.
  • Experimental result: The experiments on VisDrone2019 report excellent detection accuracy for the proposed YOLOv5 model.The supplied result passage does not provide a numerical accuracy value.

A. Experimental Setting

The experiments use the VisDrone2019 benchmark, whose drone imagery spans varied environments, object densities, weather, and lighting conditions. The setup includes center cropping and standard training, anchor, and augmentation configurations.

  • Dataset: VisDrone2019 contains 10,209 static images and 261,908 frames from 288 video clips captured by diverse drone-mounted cameras.The dataset covers urban and rural environments, sparse to crowded scenes, and varied weather and lighting conditions.
  • Dataset: VisDrone2019 is split into 6,471 training, 548 validation, and 1,610 testing images.These are the dataset partitions used in the experiment.
  • Dataset: The dataset includes 10 object classes, including pedestrians, vehicles, bicycles, buses, and motorcycles.The listed classes also include people, vans, trucks, tricycles, and awning-tricycles.
  • Data augmentation: Many objects are located near image centers, motivating center-crop augmentation during the experiment.The center crop uses half the original image height and width and was observed to improve overall model performance.
  • Training settings: Training uses 640 × 640 inputs, batch size 128, 300 epochs, Adam with an initial learning rate of 0.001, and early stopping patience 15.The supplied implementation passage also states that objectness, box, and class loss weights are 0.5, 0.05, and 0.25, respectively.
  • Data augmentation: The experiment uses Mosaic, Copy paste, Random affine, MixUp, HSV augmentation, and Cutout among its data-augmentation techniques.Center crop is additionally adopted.

C. Evaluation criterion

Object detection performance is evaluated using IoU, Precision, Recall, and mAP, with mAP reported at specified IoU thresholds.

  • IoU, Precision, Recall, and mAP are common criteria for evaluating object detection algorithms.

2) Precision:

Precision measures correctness among positive predictions, while Recall measures the proportion of actual positives predicted correctly.

  • Precision is the proportion of correctly predicted samples among all positive predictions.Precision = True positives / (True positives + False positives).
  • Recall is the proportion of actual positive samples predicted correctly.Recall = True positives / (True positives + False negatives).

4) mAP:

mAP summarizes average precision across classes and IoU thresholds, and the experiments compare HIC-YOLOv5 with YOLOv5s and other detection models on VisDrone.

  • mAP is the mean Average Precision across all classes; mAP@0.5 uses IoU 0.5, while mAP@[.5:.95] averages IoU values from 0.5 to 0.95.
  • The detection comparison between YOLOv5s and HIC-YOLOv5 is shown in Fig. 7.
  • HIC-YOLOv5 was compared with other detection models evaluated on VisDrone2019.The paper notes that mAP is incomparable with because that method used a different dataset.

E. Ablation Study

Ablation experiments examine the additional prediction head, involution block, and CBAM, with the fourth prediction head providing the strongest reported contribution.

  • The fourth prediction head improved mAP@.5 by 8.31% and mAP@[.5:.95] by 5.51% in the ablation study.
  • CBAM and Involution alone could not improve the model without the fourth prediction head.
  • The conclusion attributes the improved small-object detection performance to higher-resolution feature maps, increased channel information, CBAM attention, and center-crop augmentation.
Loading 2309.16393v2…