ECF-YOLOv7-Tiny: Improving Feature Fusion and the Receptive Field for Lightweight Object Detectors

Dan-Sebastian Bacea, Florin OnigaWACV 2025

TL;DR

This paper introduces ECF-YOLOv7-tiny, a novel lightweight object detection architecture that enhances feature fusion and receptive field capabilities. By integrating new modules like CSE and CSP-CAM and applying model scaling techniques, it achieves state-of-the-art performance on resource-constrained devices without sacrificing real-time inference speeds.

What Problem Does This Paper Address?

Real-time object detection is crucial for many applications, especially those deployed on edge devices with limited computational power (e.g., NVIDIA Jetson Nano, ARM boards). While existing lightweight object detectors like YOLOv7-tiny offer good speed, they often lag behind larger models in detection accuracy. This gap arises from limitations in how these models fuse features from different scales and how effectively they capture contextual information within images. The challenge is to improve both accuracy and feature representation without significantly increasing computational cost, making them suitable for practical, low-cost deployments.

Methods

The authors propose ECF-YOLOv7-tiny, built upon the YOLOv7-tiny architecture, by introducing several key improvements:

  • Convolutional Squeeze-and-Excitation (CSE) Module: This module enhances feature fusion by incorporating channel attention. Unlike traditional SE blocks that use fully-connected layers, CSE employs 1x1 convolutional layers for faster computation and adds a subsequent convolutional layer for better feature integration.
  • Cross-Stage Partial Context Augmentation Module (CSP-CAM): To expand the receptive field and improve feature extraction, CSP-CAM is integrated into the backbone. It branches the input into multiple paths with dilated convolutions at different rates, allowing the model to capture broader contextual information. This module is inspired by existing context augmentation techniques but optimized for efficiency.
  • Mish Activation Function: The authors replace the Leaky ReLU activation function with Mish, which has been shown to improve convergence and detection performance with minimal impact on inference speed.
  • Model Scaling: To further optimize for real-time performance, they apply model scaling techniques. Specifically, they reduce the number of convolutional filters in the layers preceding the detection heads by 50% and increase the squeeze factor in the CSE blocks. This reduces model size and computational load without a significant drop in accuracy.

Datasets

  • MS COCO 2017: This dataset was used for training and evaluating the proposed ECF-YOLOv7-tiny model. It contains 80 object categories and is a standard benchmark for object detection tasks. The authors used the 2017 splitting and evaluated on the validation set.

Key Insights and Ideas

  • Feature Fusion and Receptive Field are Key Bottlenecks: The paper convincingly demonstrates that improving how features are fused across different scales and how effectively the model captures context (receptive field) are critical for boosting the performance of lightweight detectors.
  • Efficiency Through Architectural Tweaks: The CSE and CSP-CAM modules, along with the replacement of fully-connected layers with convolutions in CSE, highlight that significant performance gains can be achieved with minimal computational overhead.
  • Model Scaling for Real-Time Deployment: The study shows that strategic model scaling, particularly reducing filters in later layers and adjusting squeeze factors, is a viable approach to meet strict real-time inference requirements on edge devices without drastically compromising accuracy.
  • Generalization of Novel Modules: The successful integration of CSE and CSP-CAM into another lightweight detector (YOLOv9-t) underscores the broad applicability and effectiveness of these proposed modules.

Key Results

  • ECF-YOLOv7-tiny: Achieves 37.8% mAP@[0.5:0.95] on the MS COCO test set at 9.3 FPS on an NVIDIA Jetson Nano, outperforming other state-of-the-art lightweight detectors at a 416 input resolution.
  • Mish Activation: Improved AP95 by 1.2% and APal by 1.5% compared to Leaky ReLU in YOLOv7-tiny.
  • CSE Module: Increased AP95 by 0.8% and APal by 0.4% over baseline YOLOv7-tiny with SE blocks, despite a slight increase in parameters and BFLOPs.
  • CSP-CAM Module: Led to a significant improvement in detection performance, increasing AP95 by 2.6% and APal by 3.2% compared to the baseline with CSE.
  • Model Scaling (ECF-YOLOv7-tiny-s): Reduced parameters by 21.7% and BFLOPs by 9.7% with a squeeze factor of 16 in CSE blocks, while maintaining competitive accuracy and achieving 9.3 FPS on Jetson Nano.
  • Comparison with YOLOv9-t: ECF-YOLOv7-tiny-s achieves 2.2 FPS more than YOLOv9-t on Jetson Nano, demonstrating better efficiency on edge devices.

Limitations and Future Work

  • Increased Parameters and BFLOPs: While the proposed modules improve accuracy, they do introduce a slight increase in parameters and BFLOPs compared to the original YOLOv7-tiny. The model scaling techniques help mitigate this, but it remains a trade-off.
  • Real-time Performance Trade-off: The ECF-YOLOv7-tiny achieves real-time performance but is slower than the original YOLOv7-tiny on the Jetson Nano (8.8 FPS vs. 14.4 FPS). The scaled version (ECF-YOLOv7-tiny-s) improves this to 9.3 FPS.
  • Future Work: The authors suggest further exploration of model scaling techniques and potentially applying their modules to other architectures to further push the boundaries of lightweight object detection. They also mention that their design choices were guided by experiments and ablation studies, implying that further fine-tuning or exploring different combinations could yield additional improvements.

Summary updated

BibTeX
@inproceedings{bacea2025ecf,
  author = {Dan-Sebastian Bacea and Florin Oniga},
  title = {{ECF-YOLOv7-Tiny: Improving Feature Fusion and the Receptive Field for Lightweight Object Detectors}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure