Bottleneck Transformers for Visual Recognition

27 Jan 2021  ·  Aravind Srinivas, Tsung-Yi Lin, Niki Parmar, Jonathon Shlens, Pieter Abbeel, Ashish Vaswani ·

We present BoTNet, a conceptually simple yet powerful backbone architecture that incorporates self-attention for multiple computer vision tasks including image classification, object detection and instance segmentation. By just replacing the spatial convolutions with global self-attention in the final three bottleneck blocks of a ResNet and no other changes, our approach improves upon the baselines significantly on instance segmentation and object detection while also reducing the parameters, with minimal overhead in latency... Through the design of BoTNet, we also point out how ResNet bottleneck blocks with self-attention can be viewed as Transformer blocks. Without any bells and whistles, BoTNet achieves 44.4% Mask AP and 49.7% Box AP on the COCO Instance Segmentation benchmark using the Mask R-CNN framework; surpassing the previous best published single model and single scale results of ResNeSt evaluated on the COCO validation set. Finally, we present a simple adaptation of the BoTNet design for image classification, resulting in models that achieve a strong performance of 84.7% top-1 accuracy on the ImageNet benchmark while being up to 2.33x faster in compute time than the popular EfficientNet models on TPU-v3 hardware. We hope our simple and effective approach will serve as a strong baseline for future research in self-attention models for vision. read more

PDF Abstract

Datasets


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Object Detection COCO minival BoTNet 152 (Mask R-CNN, single scale, 72 epochs) box AP 49.5 # 15
AP50 71 # 5
AP75 54.2 # 11
Instance Segmentation COCO minival BoTNet 152 (Mask R-CNN, single scale, 72 epochs) mask AP 43.7 # 14
Object Detection COCO minival BoTNet 200 (Mask R-CNN, single scale, 72 epochs) box AP 49.7 # 14
AP50 71.3 # 4
AP75 54.6 # 10
Object Detection COCO minival BoTNet 50 (72 epochs) box AP 45.9 # 28
Instance Segmentation COCO minival BoTNet 50 (72 epochs) mask AP 40.7 # 21
Instance Segmentation COCO minival BoTNet 200 (Mask R-CNN, single scale, 72 epochs) mask AP 44.4 # 12
Image Classification ImageNet ResNet-101 Top 1 Accuracy 80% # 221
Top 5 Accuracy 95% # 98
Number of params 44.4M # 113
Image Classification ImageNet BoTNet T7 Top 1 Accuracy 84.7% # 69
Top 5 Accuracy 97% # 33
Number of params 75.1M # 69
Image Classification ImageNet BoTNet T3 Top 1 Accuracy 81.7% # 177
Top 5 Accuracy 95.8% # 71
Number of params 33.5M # 138
Image Classification ImageNet SENet-101 Top 1 Accuracy 81.4% # 189
Top 5 Accuracy 95.7% # 77
Number of params 49.2M # 107
Image Classification ImageNet SENet-350 Top 1 Accuracy 83.8% # 100
Top 5 Accuracy 96.6% # 48
Image Classification ImageNet BoTNet T6 Top 1 Accuracy 84% # 87
Top 5 Accuracy 96.7% # 44
Number of params 53.9M # 104
Image Classification ImageNet BoTNet T5 Top 1 Accuracy 83.5% # 109
Top 5 Accuracy 96.5% # 51
Number of params 75.1M # 69
Image Classification ImageNet BoTNet T4 Top 1 Accuracy 82.8% # 133
Top 5 Accuracy 96.3% # 58
Number of params 54.7M # 103
Image Classification ImageNet ResNet-50 Top 1 Accuracy 78.8% # 264
Top 5 Accuracy 94.5% # 119
Number of params 25.5M # 156
Image Classification ImageNet BoTNet T7-320 Top 1 Accuracy 84.2% # 78
Top 5 Accuracy 96.9% # 37
Number of params 75.1M # 69
Image Classification ImageNet SENet-152 Top 1 Accuracy 82.2% # 159
Top 5 Accuracy 95.9% # 69
Number of params 66.6M # 80
Image Classification ImageNet SENet-50 Top 1 Accuracy 79.4% # 238
Top 5 Accuracy 94.6% # 113
Number of params 28.02M # 146

Methods