DCAN: Improving Temporal Action Detection via Dual Context Aggregation

Authors: Guo Chen, Yin-Dong Zheng, Limin Wang, Tong Lu248-257

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on Activity Net v1.3 and THUMOS-14. DCAN obtains an average m AP of 35.39% on Activity Net v1.3 and reaches m AP 54.1% at Io U@0.5 on THUMOS-14, which demonstrates DCAN can generate high-quality proposals and achieve state-of-the-art performance.
Researcher Affiliation Academia State Key Lab for Novel Software Technology, Nanjing University, China {chenguo1177, ydzheng0331}@gmail.com, {lmwang, lutong}@nju.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes We release the code at https://github.com/cg1177/DCAN.
Open Datasets Yes We conduct extensive experiments on Activity Net v1.3 and THUMOS-14. (Jiang et al. 2014) (Heilbron et al. 2015)
Dataset Splits Yes Activity Net v1.3. (Heilbron et al. 2015) is a large-scale action understanding dataset, which consists of 19,994 videos for training, 4,728 for validation, and 5,044 for testing, with 200 action classes.
Hardware Specification Yes Inference speed here is the seconds (s) cost Ttotal for processing a 3-minute video using an Nvidia 1080-Ti graphics card.
Software Dependencies No The paper mentions using a 'two-stream network' and 'TSN' but does not specify software libraries with version numbers (e.g., PyTorch, TensorFlow, CUDA, specific Python packages).
Experiment Setup Yes For Activity Net v1.3, [...] We set the batch size to 16 and the learning rate to 0.001 for the first 7 epochs and 0.0001 for the following 3 epochs. For THUMOS-14, [...] We set the batch size to 16 and the learning rate to 0.0004 for all 5 epochs. The Nb is set to 6 on Activity Net v1.3 and 7 on THUMOS-14. The Nbase, Nsample, rsmooth and G are set to 3, 32, 3 and 2. In the post-processing, the Soft-NMS threshold is set to 0.5 to pick the top Nfinal confident predictions, where Nfinal is 100 for Activity Net v1.3 and 200 for THUMOS-14. [...] γ is a hyperparameter for adjusting the compatibility of boundary scores and matching scores and is set as 1.5 on THUMOS-14 and 0.8 on Activity Net v1.3, respectively.