DETR ResNet-50 DC5

DETR ResNet-50 DC5 (Dilated C5 stage) modifies the standard DETR architecture by using dilated convolutions in the final stage of ResNet-50. This increases the spatial resolution of feature maps, providing better localization and significantly improving small object detection while maintaining the transformer-based detection approach.

When to Use DETR ResNet-50 DC5

Use DETR ResNet-50 DC5 when you have:

• Many small objects in your dataset (<32x32 pixels)
• Need better localization accuracy than standard DETR
• Sufficient GPU memory (higher resolution features require more memory)
• Scenarios where standard DETR misses small objects

Strengths

• Better small object detection than standard DETR (3-5% mAP improvement on small objects)
• Higher resolution features through dilated convolutions
• Improved localization with finer spatial information
• Same transformer architecture benefits as standard DETR

Weaknesses

• Higher memory usage than standard DETR ResNet-50
• Slower training due to higher resolution processing
• Slower inference (1.5-2x slower than standard DETR)
• May overfit on small datasets due to increased capacity

Parameters

Training Configuration

Training Images: Folder with images Annotations: COCO-format JSON Batch Size (Default: 2) - Range: 1-4 (memory-intensive) Epochs (Default: 1) - Range: 1-8 Learning Rate (Default: 5e-5) Eval Steps (Default: 1)

Configuration Tips

• Use when standard DETR ResNet-50 misses small objects
• batch_size=2 typical even with 16GB GPU (higher memory usage)
• Particularly effective for small object classes
• Monitor memory usage closely

Expected Performance

• 3-5% better mAP on small objects vs standard DETR ResNet-50
• Overall mAP typically 1-2% better
• Trade-off: slower inference for better small object detection

Comparison with Alternatives

vs Standard DETR ResNet-50: Choose DC5 when small objects critical, standard for faster inference

vs Deformable DETR: Deformable generally better overall, but DC5 simpler architecture if you want standard DETR with better small object handling