D-FINE vs YOLOS

D-FINE is a real-time object detection model introduced in October 2024 by researchers at the University of Science and Technology of China. It builds on the DETR family of transformer-based detectors by reformulating bounding box regression as a Fine-grained Distribution Refinement task. Rather than predicting box coordinates directly, D-FINE iteratively refines probability distributions over coordinate offsets across decoder layers, which provides finer localization granularity without adding inference cost. The architecture also replaces the encoder's CSP blocks with GELAN modules and inserts a Target Gating Layer after the decoder's cross-attention to reduce representational entanglement across queries. A second contribution, Global Optimal Localization Self-Distillation, transfers localization knowledge from refined deeper-layer predictions back to earlier decoder layers through internal self-distillation.

D-FINE is released in five model sizes (Nano, Small, Medium, Large, and X), with D-FINE-L achieving 54.0% AP on the Microsoft COCO benchmark at 124 FPS on an NVIDIA T4 GPU, and D-FINE-X reaching 55.8% AP at 78 FPS. Pretraining on the Objects365 dataset further improves accuracy to 57.1% AP for the L variant and 59.3% AP for the X variant. The paper was accepted at ICLR 2025 as a Spotlight. Code and pretrained weights are released under the Apache 2.0 license, making the model suitable for commercial use.

YOLOS (You Only Look at One Sequence) is a transformer-based object detection model widely distributed through Hugging Face Transformers, released in June 2021 under the MIT license. It applies a minimally adapted Vision Transformer to object detection by representing both the image and detection tokens as a flat sequence processed by standard multi-head self-attention, without convolutional components or feature pyramid networks. The architecture demonstrates that detection can be performed without region proposals or multi-scale feature fusion.

YOLOS achieves moderate performance on COCO relative to purpose-built detectors, with its primary contribution being a demonstration of the transferability of ViT pre-training to detection tasks. It is most appropriate for research contexts exploring transformer-based detection architectures and for scenarios where architectural simplicity is preferred over peak accuracy.