Welcome to the repository dedicated to finetuning DETR on the SKU110K dataset! 🚀
I am excited to share that our trained checkpoint, DETR-Resnet-50 configured for SKU110K with 400 queries, is now available on HuggingFace. 🎉
Finetuning DETR with a num_queries parameter different from the default (100) is challenging. Our experiments show that without a proper initialization strategy, training tends to fail — resulting in low mean Average Precision (mAP) even after extensive training.
I discovered that using pretrained weights to initialize num_queries significantly improves performance. It leverages the pretrained model's ability to detect objects across various image areas, making a small adjustment to specialize in the new dataset much easier than starting from scratch.
🔍 Explore the load_pretrained_num_queries function in detr_model.py to see how we implement this strategy.
When num_queries is different than 100, we've found that duplicating the pretrained num_queries=100 weights and introducing minor noise sets the stage for success. So for num_queries=400 we take num_queries=100 and duplicate them 4 times.
After extensive experimentation with various configurations, this approach stood out, achieving a 59.0 mAP on the SKU110K validation set.