[Paper Review] Cattle Detection Occlusion Problem

Aparna Mendu, Bhavya Sehgal+1 more|arXiv (Cornell University)|Dec 21, 2022

Food Supply Chain Traceability4 citations

TL;DR

This paper proposes YOLOv7 as the most effective deep learning model for detecting occluded and small cattle in drone-captured RGB and thermal images, outperforming RetinaNet (with ResNet50 and EfficientNet backbones) and Mask R-CNN. Using image tiling (2x1 and 3x3) to enhance small object detection, YOLOv7 achieved the highest precision of 61.2%, demonstrating superior performance in challenging occlusion scenarios.

ABSTRACT

The management of cattle over a huge area is still a challenging problem in the farming sector. With evolution in technology, Unmanned aerial vehicles (UAVs) with consumer level digital cameras are becoming a popular alternative to manual animal censuses for livestock estimation since they are less risky and expensive.This paper evaluated and compared the cutting-edge object detection algorithms, YOLOv7,RetinaNet with ResNet50 backbone, RetinaNet with EfficientNet and mask RCNN. It aims to improve the occlusion problem that is to detect hidden cattle from a huge dataset captured by drones using deep learning algorithms for accurate cattle detection. Experimental results showed YOLOv7 was superior with precision of 0.612 when compared to the other two algorithms. The proposed method proved superior to the usual competing algorithms for cow face detection, especially in very difficult cases.

Motivation & Objective

• Address the persistent challenge of detecting occluded and small cattle in large-scale drone-based livestock monitoring.
• Evaluate and compare state-of-the-art object detection models—YOLOv7, RetinaNet (with ResNet50 and EfficientNet backbones), and Mask R-CNN—for cattle detection under occlusion.
• Improve detection accuracy for hidden or partially visible cattle using image tiling techniques (2x1 and 3x3) to enhance resolution and feature learning.
• Demonstrate the effectiveness of YOLOv7’s Extended-ELAN and coarse-to-fine label assignment in handling class imbalance and improving detection in complex aerial scenes.
• Establish a benchmark for cattle detection in real-world drone datasets, particularly under challenging occlusion and small-object conditions.

Proposed method

• Employed YOLOv7 with Extended-ELAN (E-ELAN) architecture to enhance feature learning and gradient flow, improving detection accuracy in occluded scenarios.
• Applied image tiling (2x1 and 3x3) to divide high-resolution drone images into smaller patches, enabling better detection of small and hidden cattle.
• Trained RetinaNet with ResNet50 and EfficientNet backbones using focal loss to mitigate class imbalance between foreground and background objects.
• Used Mask R-CNN with ResNet50 backbone for instance segmentation, adding a mask head to predict pixel-level object boundaries for improved localization.
• Implemented coarse-to-fine label assignment in YOLOv7 to optimize training for multi-output detection heads and improve convergence.
• Utilized a multi-source dataset: 1161 drone-captured images (RGB and thermal) from Canchim farm, Brazil, and an additional 249-image RGB dataset from Harvard University.

Experimental results

Research questions

• RQ1 How does YOLOv7 compare to RetinaNet and Mask R-CNN in detecting occluded cattle in drone-captured aerial images?
• RQ2 What is the impact of image tiling (2x1 vs. 3x3) on the detection performance of small and hidden cattle?
• RQ3 How effective is the use of EfficientNet as a backbone in RetinaNet for detecting small livestock objects compared to ResNet50?
• RQ4 Can YOLOv7’s E-ELAN and label assignment mechanisms improve detection accuracy in high-occlusion scenarios?
• RQ5 Which model achieves the highest precision and recall for cattle detection across diverse datasets and tiling configurations?

Key findings

• YOLOv7 achieved the highest precision of 61.2% on the drone dataset using 2x1 tiling, outperforming all other models.
• RetinaNet with ResNet50 backbone achieved 58.3% precision, demonstrating strong performance but falling short of YOLOv7’s accuracy.
• RetinaNet with EfficientNet backbone achieved 50.4% precision, indicating lower effectiveness for small object detection in this context.
• Mask R-CNN performed poorly on the drone dataset but showed improved results on the external RGB dataset, achieving better performance than other models in that specific case.
• 2x1 tiling yielded better detection results than 3x3 tiling across all models, particularly enhancing small object detection in occluded scenes.
• The combination of YOLOv7 and 2x1 tiling proved most effective for detecting hidden or partially visible cattle, highlighting the importance of both model architecture and preprocessing strategy.