Reproduce “Real-time Transformer-based Open-Vocabulary Detection with Efficient Fusion Head” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tiancheng Zhao, Peng Liu, Xuan He, Lu Zhang, Kyusong Lee
OmDet-Turbo introduces the first real-time transformer-based open-vocabulary object detection method, achieving over 100 FPS with TensorRT while establishing new state-of-the-art performance on various zero-shot OVD benchmarks like ODinW and OVDEval. The model addresses efficiency limitations of prior transformer-based OVD systems through an Efficient Fusion Head and strategic architectural optimizations.
Key insights
- Real-time performance in transformer-based OVD is achievable by strategically optimizing or replacing computationally heavy components, particularly in the encoder and detection head, without sacrificing generalization.
- Decoupling prompt and label encoding, combined with a language cache, significantly enhances flexibility for large vocabularies and reduces text backbone inference time to sub-milliseconds.
- Multitask pre-training on diverse vision-language tasks, unified into a VQA-like format, is crucial for improving zero-shot generalization and adaptability to various real-world scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.