End-to-End Object Detection with Transformers by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko (2020)

This note was initially drafted with LLM assistance. Generated notes are periodically reviewed and revised by the author.
tags
Transformers, Attention, Computer vision, Object recognition, Positional encoding
source
(Carion et al. 2020)

Summary

DETR (DEtection TRansformer) reframes object detection as a direct set prediction problem, eliminating hand-designed components that traditional detectors rely on: anchor generation, non-maximum suppression (NMS), and coordinate-regression heuristics against proposals. The model is a CNN backbone (ResNet-50 / ResNet-101) feeding a standard Transformer encoder-decoder. The decoder attends to the encoder output via a small fixed set of \(N\) learned object queries (with \(N\) much larger than the number of objects in any image) and, in parallel, produces \(N\) (class, box) predictions through a shared feed-forward head. Predictions with no matched ground-truth are labeled as a special \(\varnothing\) “no object” class.

Training uses a set-based global loss built around the Hungarian algorithm: ground-truth objects and predicted slots are matched one-to-one by minimizing a pair-wise cost combining class probability and a box loss (L1 plus generalized IoU). Because the matching is unique, duplicates are suppressed by construction rather than by a post-hoc NMS step. Auxiliary Hungarian losses are added after every decoder layer to stabilize training.

On COCO, DETR matches Faster R-CNN in average precision and is notably stronger on large objects (thanks to the transformer’s global attention), while being weaker on small objects and requiring a much longer training schedule. The same model extends naturally to panoptic segmentation by adding a small mask head on top of attention maps, outperforming the baselines of the time.

Key Ideas

  • Object detection as set prediction: output a fixed-size set of tuples (class, bounding-box) in one forward pass, with \(\varnothing\) padding.
  • Bipartite matching loss (Hungarian algorithm): enforces a unique assignment between predictions and ground truth, making the loss invariant to prediction permutations and removing the need for NMS.
  • Object queries: \(N\) learned positional embeddings used as decoder input; each query specializes during training to a rough spatial region / size pattern and becomes the slot that a specific object “lands in”.
  • Parallel (non-autoregressive) decoding: all \(N\) outputs are produced in one decoder pass, a departure from autoregressive seq2seq decoders.
  • Box loss: linear combination of \(\ell_1\) and generalized IoU, to be scale-invariant.
  • Auxiliary losses: Hungarian loss applied after every decoder layer, sharing prediction heads, improves training.
  • Trivially extensible: a small segmentation head on top of the encoder attention maps turns DETR into a competitive panoptic segmentation model.
  • Conceptual simplicity: ~50 lines of PyTorch for inference; no custom CUDA ops, no specialized detection library.

Comments

DETR is an influential bridge between the transformer revolution in NLP and dense prediction tasks in computer vision. Its value is less about raw COCO numbers — it only matches a mature Faster R-CNN baseline — and more about showing that the entire hand-engineered pipeline of region proposals, anchors, and NMS can be replaced by a uniform attention-based model trained with a permutation-invariant set loss. In that sense it sits next to Vision transformer as one of the early “transformers can do vision too” works, but with a very different output structure (sets of typed elements, not a sequence of patches with a single label).

Limitations acknowledged in the paper remain instructive: small-object performance is poor (later remedied by Deformable-DETR and multi-scale variants), training converges much more slowly than RCNN-style models (500 epochs vs. ~40), and the matching becomes a bottleneck when \(N\) is large. The object-query design also foreshadows the more general pattern of learned slots attending to image features that shows up throughout modern perception stacks and in Vision Language Models.

Pleasingly, there is now a direct descendent authored by the same first author in the knowledge base — SAM 3 — which shares the lineage of transformer-based unified detection / segmentation with concept prompts.

Connections

  • Closely related to Vision transformer: both adapt standard encoder-decoder transformers to vision, but DETR predates ViT and uses a CNN backbone plus flattened feature map as the token sequence.
  • Uses the core machinery of Transformers and Attention (multi-head self- and cross-attention) as-is — the novelty is in the loss and output parameterization, not the architecture.
  • Relies on Positional encoding added at every attention layer (fixed sinusoidal for image features, learned for object queries).
  • Subsumed into the Object recognition topic as one of the first fully end-to-end detectors.
  • Direct predecessor of the line of work culminating in (Carion et al. 2025) (SAM 3), by the same first author — same transformer encoder-decoder + query-based decoding philosophy, scaled to prompt-conditioned concept segmentation.
  • Contrasts with hierarchical CV transformers like the Swin Transformer: DETR keeps a single-scale feature map and relies on the decoder to handle multi-object output, where Swin instead changes the encoder to produce multi-scale features.

Bibliography

  1. . . "End-to-End Object Detection with Transformers". https://arxiv.org/abs/2005.12872.
  2. . . "SAM 3: Segment Anything with Concepts". https://arxiv.org/abs/2511.16719. See notes

Links to this note

Last changed | authored by

Comments

Loading comments...

Leave a comment

Back to Notes