June 2020
tl;dr: Transformer used for object detection as direct set prediction .
Formulate the object detection problem as direct set prediction problem. No need for engineering-heavy anchor boxes and NMS.
The attention mechanism from transformers is similar to Non-local Networks. The attention has perfect memory and has same "distance" between any two points in the image.
Previous methods such as anchor-based or anchor-free, implicitly enables a sorting of GT and prediction. The Hungarian loss used in DETR eliminates that altogether by comparing loss between two unordered sets.
This paper is extended by Deformable DETR to speed up the training of transformers by more than x10.
- Training:
- Two step process. Bipartite matching loss + Prediction loss. (the second term is called "Hungarian loss" which is quite confusing)
- The bipartite matching forces a 1-to-1 matching, without missing.
- Null padding of training set to N.
- DETR infers a fixed size set of N predictions. Predicts normalized center coordinates wrt the input image.
- Resizing feature map HxWxd to dx(HW), as a sequence of feature dim d and length of HW.
- Decoding output is non-autoregressive parallel decoding (feed previous output to the decoder to get next output).
- Input sequence to decoder are all zero, and with learned Positional embeddings (object query, learned specialized workers attending to different types of boxes).
- Object query can be learned with SGD. It is part of the model's weight. It is a bit like PE but not exactly the same. It is essentially training different annotators to pay different part to the image and focus on certain type of bboxes.
- The N=100 decoders in a way can be seen as 100 adaptive anchors.
- Need extremely long training (300 epochs) to converge, vs 1x = 12 epochs for Faster RCNN. --> This is improved by Deformable DETR.
- Input is a flattened feature map, with shape HW x C.
- Matching loss + Hungarian loss
$L_{matching} = -\mathbb{1}(c_i \neq \varnothing)p(c_i) + \mathbb{1}(c_i \neq \varnothing) L(y_{bbox}, \hat{y}_{bbox})$ $L_{Hungarian} = -p(c_i) + \mathbb{1}(c_i \neq \varnothing) L(y_{bbox}, \hat{y}_{bbox})$ - Note that the class mask in Hungarian loss is gone. This is needed to suppress false positive or near-duplicates.
- Generalized IoU loss for scale invariant and more robust alternative to IoU or Dice loss.
- Decoder computational complexity is much smaller than encoder as num of objects N << HW.
- Seems like object detectors usually do not use Adam nor dropout. ("We are not aware of successful applications of ...")
- Key findings for explainability of DETR:
- Encoder is able to separate individual objects.
- Decoder typically attends to object extremities.
- The visualization of bbox prediction of various output slots in DETR demonstrates that each slot attends to a particular type of objects. This is highly intriguing, similar to the prototype masks in YOLACT.
- Uses updated AdamW for training.