What question did this study set out to answer?

To determine whether pure Vision Transformer architectures can effectively serve as backbone feature extractors for complex object detection tasks beyond image classification.

December 17, 2020Open Access

Toward Transformer-Based Object Detection

View Full Paper

Ask AI

Bookmark

Authors

JBJosh BealYale University EKEric KimUniversity of British Columbia Hospital ETEric TzengUC San Diego Health System

Discussion

Member takes

Implication

Computational benchmark demonstrates competitive object detection using Vision Transformers on COCO, indicating strong generalization and reduced post-processing reliance.

Key Points

To determine whether pure Vision Transformer architectures can effectively serve as backbone feature extractors for complex object detection tasks beyond image classification.
Coupled a Vision Transformer backbone with a standard Faster R-CNN detection head to create the ViT-FRCNN model.
Pretrained the transformer backbone on large image datasets and fine-tuned the complete architecture on the standard COCO benchmark.
Evaluated detection quality across input resolutions, out-of-domain datasets, varied object scales, and non-maximum suppression dependencies.
ViT-FRCNN achieved competitive detection performance on the COCO benchmark while exhibiting high pretraining capacity and fast fine-tuning.
The architecture demonstrated superior transfer performance on out-of-domain images and enhanced accuracy on large object instances relative to standard backbones.
The model exhibited a reduced reliance on standard non-maximum suppression heuristics for bounding box deduplication.

Cite This Study

Beal et al. (2020) studied this question.

synapsesocial.com/papers/6a16ae87f3be5e880d6b660d https://doi.org/10.48550/arxiv.2012.09958

View Full Paper

Ask AI

Bookmark

Implication

Computational benchmark demonstrates competitive object detection using Vision Transformers on COCO, indicating strong generalization and reduced post-processing reliance.

Toward Transformer-Based Object Detection

Discussion

Member takes

Implication

Key Points

Cite This Study

Toward Transformer-Based Object Detection

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider