SSH-YOLO Ablation Study Results for Object Detection in Road Scenes

Ablation study results for the SSH-YOLO object detection model, proposed by Tenglong Ma and last updated in April 2026. The tabular data, 5.5 KB in size, contains performance metrics from experiments on a self-built RoadScene-Complex dataset and public datasets including BDD100K, KITTI, COCO, and CityPersons. The study evaluates improvements for detecting dense, small, and occluded targets in complex traffic scenarios.

Use Cases

• Benchmarking object detection model performance on small targets (< 32x32 pixels) based on the described four-level detection system.
• Evaluating model accuracy in occluded scenes based on the spatial and channel collaborative attention module (SCSA) designed for such targets.
• Comparing inference speed and accuracy trade-offs for real-time on-board detection based on the reported 60 FPS performance.
• Analyzing the impact of architectural components like the SPDConv module on feature retention for low-resolution images.

Strengths

• Performance metrics are reported against multiple established public datasets (BDD100K, KITTI, COCO, CityPersons).
• Specific [email protected] improvements are quantified, such as a 12.4% increase over YOLOv8n on the RoadScene-Complex dataset.
• The model maintains a reported inference speed of 60 FPS, indicating a focus on real-time application.

Limitations

• Row count and specific column definitions are unknown, limiting assessment of data granularity and structure.
• The dataset is very small (5.5 KB), suggesting it contains summary results rather than raw experimental data.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

Tenglong Ma via figshare.

Collection Method

Likely contains experimental results from model training and evaluation.

Freshness

Last updated 2026-04-15 18:05:41.