Deeply-Supervised CNN Model for Action Recognition with Trainable Feature Aggregation

Authors: Yang Li, Kan Li, Xinxin Wang

IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on two action recognition datasets: HMDB51 and UCF101. Results show that our model outperforms the state-of-the-art methods.
Researcher Affiliation Academia Yang Li, Kan Li , Xinxin Wang School of Computer Science, Beijing Institute of Technology, Beijing, China {yanglee, likan, wangxx}@bit.edu.cn
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes We evaluate the proposed models on two popular action classification benchmarks, HMDB51 [Kuehne et al., 2011] and UCF101 [Soomro et al., 2012].
Dataset Splits Yes Both datasets are split into three parts, and we report final performance averaged over all three splits.
Hardware Specification Yes We train it on a PC with 3.4GHz CPU, a TITAN X GPU, and 64G RAM.
Software Dependencies No The paper only mentions software names (Inception with Batch Normalization, Adam) but does not provide specific version numbers for key libraries or frameworks.
Experiment Setup Yes Thus, we finally selected T = 10, which is a good trade-off between accuracy and efficiency. For the parameters of the aggregation module, we experiment with different values of K (K = 16, 32 and 64) and obtain the best performance when K = 32. The network is trained with a momentum of 0.9 and a weight of decay of 4e 5. learned by back-propagating the loss function defined in Eq. 8 using the Adam [Kingma and Ba, 2014] with ϵ = 1e 4.