Transformers have established their importance in natural language processing (NLP), their ability to learn a general language representation and transfer it to specific tasks has made them invaluable. And with the advent of Vision Transformer (ViT) we saw that standard Transformer encoder architecture inherited from NLP can perform surprisingly well on image recognition at scale. Pre-trained transformers can be fine-tuned on sentence-level tasks, as well as more complex tasks that involve producing fine-grained output at the token level. This raises a question: Can vision transformers like ViT transfer to more complex computer vision tasks such as object detection?
ViT-RCNN already leveraged a pre-trained ViT as the backbone for an R-CNN object detector but it still relies heavily on convolution neural networks and strong 2D inductive biases. Briefly put, ViTFRCNN interprets the output sequence of ViT to 2D spatial feature maps and utilizes region-wise pooling operations and R-CNN architecture to decode the features for object-level perception. There have been other similar works, like DEtection TRansformer (DETR), that introduce 2D inductive bias by leveraging pyramidal feature hierarchies and CNNs.
However, these architectures are performance-oriented and they don’t reflect the properties of the vanilla Transformer. ViT is designed to model long-range dependencies and global contextual information instead of local and region-level relations. Moreover, ViT doesn’t have hierarchical architecture like CNNs to handle the large variations in the scale of visual entities. But Transformers are born to transfer, so we can’t dismiss them without testing whether a pure ViT can transfer pre-trained general visual representations from image-level recognition to the much more complicated 2D object detection task
To test the efficacy of vanilla transformer models, Yuxin Fang, Bencheng Liao, et al, created You Only Look at One Sequence (YOLOS), a series of object detection models based on the ViT architecture with the fewest possible modifications and inductive biases.
Architecture & Approach
YOLOS closely follows the ViT architecture, there are two simple changes:
- YOLOS drops the [CLS] token used for image classification and adds one hundred randomly initialized detection [DET] tokens to the input patch embedding sequence for object detection.
- The image classification loss used in ViT is replaced with a bipartite matching loss to perform object detection similar to DETR.
YOLOS is pre-trained on the relatively small ImageNet-1k dataset. It is then fine-tuned using the COCO object detection dataset. It is important to reiterate that the whole model isn’t trained on the COCO dataset per se, YOLOS learns a general representation using ImageNet-1k and is then fine-tuned on the COCO dataset. If you’ve trained a custom object detection model, or simply employed transfer learning, you have taken a pre-trained model, frozen most of its layers, and fine-tuned the final few layers for your specific dataset/use case. Similarly, in YOLOS all the parameters are initialized with the ImageNet-1k pre-trained weights except for the MLP heads for classification & bounding box regression, and the one hundred [DET] tokens.
The randomly initialized detection [DET] tokens are used as substitutes for object representation. This is done to avoid inductive bias of 2D structure and any prior knowledge of the task that can be introduced during label assignment. When YOLOS models are fine-tuned on COCO, an optimal bipartite matching between predictions generated by [DET] tokens and the ground truth is established for each forward pass. This serves the same purpose as label assignment but is completely unaware of the input 2D structure, or even that it is 2D in nature.
What this means is that YOLOS does not need to re-interpret ViT’s output sequence to a 2D feature map for label assignment. YOLOS is designed with minimal inductive bias injection in mind. The only inductive biases it has are inherent from the patch extraction at the network stem part of ViT and the resolution adjustment for position embeddings. Besides these, YOLOS adds no non-degenerated convolutions, i.e., non 1 x 1 convolution on ViT. Any performance-oriented aspects of modern CNN architectures such as pyramidal feature hierarchy, region-wise pooling, and 2D spatial attention are not added.
This is all done in order to better demonstrate the versatility and transferability of Transformer from image recognition to object detection in a pure sequence-to-sequence manner with minimal knowledge about the spatial structure of the input. And as YOLOS doesn’t know about the spatial structure and geometry, it is feasible for it to perform any dimensional object detection as long as the input is flattened to a sequence in the same way. In addition to that, YOLOS can easily be adapted to various Transformers available in NLP and computer vision.
Results
Comparisons with Tiny-sized CNN Detectors
To test its capabilities YOLOS was compared with some modern CNN-based object detectors. The smaller YOLOS variant YOLOS-Ti achieves impressive performance compared with existing highly-optimized CNN object detectors like YOLOv4 Tiny. It has strong AP and is competitive in FLOPs and FPS even though it was not intentionally designed to optimize these factors.
Comparisons with DETR
Although YOLOS-Ti performs better than the DETR counterpart, the larger YOLOS models with width scaling are less competitive. YOLOS-S with more computations is 0.8 AP lower compared with a similar-sized DETR model. What’s even worse is that YOLOS-B cannot beat DETR with over 2× parameters and FLOPs. And despite the fact that YOLOS-S with dwr(fast) scaling outperforms the DETR counterpart, the performance gain cannot be clearly explained by the corresponding CNN scaling methodology.
With these results, we have to keep in mind that YOLOS is not designed to be yet another high-performance object detector. It is merely a touchstone for the transferability of ViT from image recognition to object detection. To compare it with state-of-the-art models like YOLOR or YOLOX would be unfair. There are still many challenges that need to be resolved, but the performance on COCO is promising nonetheless. These initial findings effectively demonstrate the versatility and generality of Transformer to downstream tasks.