End-to-End Object Detection with Transformers

Problem

Framing

Anchor- and proposal-based detectors solved detection through surrogate dense prediction plus hand-tuned post-processing. DETR replaces that stack with direct set prediction: a transformer emits a fixed detection set, and Hungarian matching enforces one-to-one supervision. On COCO, it reaches 42.0 AP with ResNet-50 while removing anchors and NMS.

Currently Used Methods

Foundational

• @renFasterRCNN2015 — two-stage detection with proposals and box/class heads.
  • Limitation in context: depends on anchors, proposal heuristics, and NMS.
• @vaswaniAttentionAllNeed2017 — encoder-decoder self-attention for global sequence modeling.
  • Limitation in context: no set loss or detection-specific object queries.
• @heMaskRCNN2017 — extends two-stage detectors to instance-level masks.
  • Limitation in context: keeps proposal pipeline and handcrafted post-processing.
• @redmonYOLO2016 — single-stage dense prediction over grid locations.
  • Limitation in context: still relies on surrogate local predictions and duplicates control.

Proposed Method

Architecture

DETR uses a CNN backbone, a transformer encoder-decoder, and $N=100$ learned object queries. The reference model uses 6 encoder layers, 6 decoder layers, width $d=256$, and 8 attention heads; each query feeds an FFN for class and normalized box prediction.

Loss / Objective

Training solves set prediction by Hungarian assignment, then applies classification plus box loss on the matched pairs.

Sampling Rule / Algorithm

Inference is one parallel decoder pass over the fixed query set, followed by per-query class and box prediction.

Training Procedure

• Optimizer: AdamW.
• Transformer learning rate: $10^{-4}$.
• Backbone learning rate: $10^{-5}$.
• Weight decay: $10^{-4}$.
• Dropout: $0.1$.
• Batch size: $64$ total, 4 images per GPU on 16 V100s.
• Base schedule: 300 epochs, LR drop after 200.
• Comparison schedule: 500 epochs, LR drop after 400.
• Gradient clipping max norm: $0.1$.
• Augmentation: scale resize to short side $[480,800]$, long side $\leq 1333$, random crop with probability $0.5$.
• Auxiliary Hungarian losses after each decoder layer.

Evaluation

Datasets

• COCO 2017 detection: 118k train, 5k val.
• COCO 2017 panoptic segmentation.

Metrics

• Bounding-box AP, AP50, AP75.
• AP by scale: $\mathrm{AP}_S$, $\mathrm{AP}_M$, $\mathrm{AP}_L$.
• Panoptic Quality: PQ, PQth, PQst.
• Throughput and compute: FPS, GFLOPS.

Headline results

• COCO detection, DETR-R50: AP 42.0, AP50 62.4, AP75 44.2.
• COCO detection, DETR-DC5-R101: AP 44.9, AP50 64.7, AP75 47.7.
• COCO detection, DETR-R50: $\mathrm{AP}_L$ 61.1, exceeding Faster R-CNN-FPN+ by 7.7.
• COCO panoptic, DETR-R101: PQ 45.1, PQth 50.5, PQst 37.0.
• COCO panoptic, DETR-DC5-R50: PQ 44.6.

Table 1: Comparison with Faster R-CNN on COCO validation.

Ablations

• Encoder depth: 0 to 12 layers raises AP from 36.7 to 41.6; large-object AP gains most.
• Decoder depth: AP and AP50 improve at every layer; final-vs-first gain is +8.2/+9.5.
• NMS: helps early decoder layers, hurts final layers; duplicate suppression is learned.
• Loss terms: GIoU alone nearly matches baseline; $L_1$ alone is weak, but improves medium/large boxes with GIoU.
• Positional encodings: removing spatial encodings drops AP from 40.6 to 32.8.

Method Strengths and Weaknesses

Strengths

• Removes anchors, proposals, and NMS with one bipartite set loss.
• Matches Faster R-CNN AP with similar parameter count: 42.0 AP at 41M parameters.
• Strong large-object detection: $\mathrm{AP}_L=61.1$ for DETR-R50.
• Extends cleanly to panoptic segmentation, reaching 45.1 PQ with DETR-R101.

Weaknesses

• Small-object performance trails tuned Faster R-CNN baselines: 20.5 APS for DETR-R50.
• Training is long: 300 epochs baseline, 500 for strongest comparisons.
• DC5 improves accuracy but roughly doubles overall compute.
• Final AP depends on auxiliary decoder losses and long optimization.

Suggestions from the authors

• Improve small-object performance, analogous to FPN gains in two-stage detectors.
• Explore stronger unified models for panoptic segmentation.
• Study better ways to combine box loss terms and match costs.
• Extend the end-to-end set-prediction design to more complex structured vision tasks.

Links

Prior Papers

• @renFasterRCNN2015 — DETR is positioned directly against Faster R-CNN and removes its proposal and NMS machinery.
• @vaswaniAttentionAllNeed2017 — DETR imports the transformer encoder-decoder and adapts it to parallel object-set prediction.
• @heMaskRCNN2017 — the panoptic extension mirrors Mask R-CNN’s task expansion while avoiding proposal-centric design.

Further Papers

• @kirillovSAM2023 — both pursue promptable or set-based segmentation from transformer-style image representations.
• @dosovitskiyViT2020 — later vision transformers strengthen the architectural shift DETR initiated for recognition pipelines.
• @caronDINO2021 — self-supervised transformer backbones become natural descendants for DETR-style detection systems.