End-to-End Object Detection with Transformers
Links
Methods
Problem
- Two-stage detector
- proposal
- single stage methods
- anchors
- grid of possible object center
- -> set based loss
- learnable NMS methods
- relation networks
Innovation
We present a new method that views object detection as a direct set prediction problem.
- object detection -> direct set prediction
- no hand-designed components like NMS, anchor generation
- DEtection TRansformer
- bipartite matching 二分图匹配 between predicted and ground-truth objects
- simplicity
advantage
model architecture
key points
- CNN → to extract features
- transformer encoder → extract global information, so that only one bounding box will be generated for one object
- object query
- encoder learn to differ objects
- decoder learn edges
loss
- fixed-size set of N predictions N=100
- optimal bipartite matching D-math
- Hungarian algorithm
- linear sum assignment
- classification loss and bounding box loss
code
complement
Information out of the article
- Deformable DETR
- Omni~
- up~
- PNP~
- SMAC~
- DAB~
- SAM~
- DN~
- OW~
- OV~