Hierarchical Online Instance Matching for Person Search
Authors: Di Chen, Shanshan Zhang, Wanli Ouyang, Jian Yang, Bernt Schiele10518-10525
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | From the experiments on two standard person search benchmarks, i.e. CUHK-SYSU and PRW, we achieve state-of-the-art performance, which justifies the effectiveness of our proposed HOIM loss on learning robust features. |
| Researcher Affiliation | Collaboration | Di Chen,1,4 Shanshan Zhang,1,3 Wanli Ouyang,5 Jian Yang,1,2 Bernt Schiele4 1PCA Lab, Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministryof Education, School of Computer Science and Engineering, Nanjing University of Science and Technology 2Jiangsu Key Lab of Image and Video Understanding for Social Security 3Science and Technology on Parallel and Distributed Processing Laboratory (PDL) 4MPI Informatics, 5The University of Sydney, Sense Time Computer Vision Research Group, Australia |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | https://github.com/Dean Chan/HOIM-PyTorch |
| Open Datasets | Yes | We follow the standard train/test split pre-defined by the dataset, where the training set consists of 11,206 images and 5,532 identities, whilst the testing set contains 2,900 probe persons and 6,978 gallery images. ... It splits 5,704 images with 482 different IDs for training and 2,057 probe persons w.r.t. 6,112 gallery images for testing. |
| Dataset Splits | No | The paper explicitly mentions train and test sets but does not detail a separate validation set or its split. |
| Hardware Specification | Yes | We train our model jointly with a batch size of 5 on a single NVIDIA Tesla P40 GPU. |
| Software Dependencies | No | The paper mentions 'PyTorch (Paszke et al. 2017)' as the base for implementation but does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | The momentum η, softmax temperature τ and importance decay factor k of HOIM are set to 0.5, 1/30 and 0.99 respectively. Sizes of the embedding buffers, i.e. N, M and B, are set individually for different datasets. For CUHK-SYSU, they are 5,532, 5,000 and 5,000; for PRW, N is set to 482, while M and B are both reduced to 500 to balance the probability distribution. We train our model jointly with a batch size of 5 on a single NVIDIA Tesla P40 GPU. The target learning rate is set to 0.003, which is gradually warmed-up at the first epoch and decayed by a factor of 0.1 at the 16th epoch. The training process converges at epoch 22. |