Pose-Invariant Face Recognition via Adaptive Angular Distillation

Authors: Zhenduo Zhang, Yongru Chen, Wenming Yang, Guijin Wang, Qingmin Liao3390-3398

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two challenging benchmarks (IJB-A and CFP-FP) show that our approach consistently outperforms the existing methods.
Researcher Affiliation Academia Zhenduo Zhang, Yongru Chen, Wenming Yang*, Guijin Wang, Qingmin Liao Shenzhen International Graduate School/Department of Electronic Engineering, Tsinghua University, China
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 use the popular dataset MS-Celeb-1M (Guo et al. 2016) for training both teacher network and the student network. For evaluation, we adopt two benchmarks for pose-invariant face recognition: CFP-FP (Sengupta et al. 2016) and IJB-A (Klare et al. 2015) datasets with official evaluation protocols (Sengupta et al. 2016; Klare et al. 2015).
Dataset Splits No The paper mentions cleaning the MS-Celeb-1M dataset and using 'official evaluation protocols' for CFP-FP and IJB-A, but it does not specify explicit validation dataset splits (e.g., percentages, counts, or specific predefined splits) for its own training process to reproduce the data partitioning.
Hardware Specification Yes We use 4 Ge Force GTX 1080 GPUs for training and we select Res Net50, Res Net34 and Res Net18 as backbones due to the limitation of computation capacity.
Software Dependencies No The paper mentions data pre-processing and model architecture but does not provide specific software details like library names with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes The initial learning rate is 0.001 and the default hyper-parameters of our method are λ1 = 0.5, λ2 = 0.5, µ1 = 0.01 and µ2 = 0.4. We set N = 20, C = 5 and M = 8. For all the models during inference stage, we extract the 512-D feature embeddings and use cosine distance as the metric.