BasicTAD: An astounding RGB-Only baseline for temporal action detection

Temporal action detection (TAD) is extensively studied in the video understanding community by generally following object pipeline images. However, complex designs are not uncommon TAD, such as two-stream feature extraction, multi-stage training, temporal modeling, and global context fusion. In this paper, we do aim to introduce any novel technique for TAD. Instead, study a simple, straightforward, yet must-known baseline given current status of design low efficiency our simple (BasicTAD), decompose TAD into several essential components: data sampling, backbone design, neck construction, head. We investigate existing techniques each component and, more importantly, perform end-to-end training over entire thanks simplicity design. As result, BasicTAD yields an astounding real-time RGB-Only very close state-of-the-art methods with inputs. addition, further improve preserving spatial information network representation (termed PlusTAD). Empirical results demonstrate that PlusTAD efficient significantly outperforms previous on datasets THUMOS14 FineAction. Meanwhile, also in-depth visualization error analysis proposed method try provide insights problem. Our approach can serve strong future research. The code model released at https://github.com/MCG-NJU/BasicTAD.