ResNet Backbone Depth and Regularization in RT-DETR Under Environmental Variations
A study evaluates RT-DETR for detecting round objects in competitive robotics, comparing four ResNet backbones (ResNet18, ResNet34, ResNet50, ResNet101) with varying dropout rates. Environmental conditions like lighting and background contrast primarily affect prediction confidence, while inference latency remains stable and classification accuracy stays high (near or above 1.00). The research addresses a gap in literature on transformer-based detectors regarding backbone scale and environmental settings.
Key facts
- RT-DETR evaluated for round object detection under environmental variations
- Four ResNet backbones compared: ResNet18, ResNet34, ResNet50, ResNet101
- Dropout rates analyzed for effect on confidence and accuracy
- Environmental conditions impact prediction confidence
- Inference latency largely unaffected by environmental changes
- Classification accuracy consistently high (near or above 1.00)
- Study addresses lack of literature on backbone scale and environmental settings in transformer detectors
- Models trained under same configuration, tested with lighting and background contrast changes
Entities
—