Hi-Fi: Hierarchical Feature Integration for Skeleton Detection
Authors: Kai Zhao, Wei Shen, Shanghua Gao, Dandan Li, Ming-Ming Cheng
IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our method significantly outperforms the state-of-the-art methods in terms of effectively fusing features from very different scales, as evidenced by a considerable performance improvement on several benchmarks.In this section, we discuss the implementation details and report the performance of the proposed method on several open benchmarks. |
| Researcher Affiliation | Academia | Kai Zhao1, Wei Shen2, Shanghua Gao1, Dandan Li2, Ming-Ming Cheng1 1 College of Computer and Control Engineering, Nankai University 2 Key Laboratory of Specialty Fiber Optics and Optical Access Networks, Shanghai University kz@mail.nankai.edu.cn, cmm@nankai.edu.cn |
| Pseudocode | No | The paper describes the approach and its formulation using text and diagrams (Fig. 3, Fig. 4), but does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at http://mmcheng.net/hifi. |
| Open Datasets | Yes | The experiments are conducted on four popular skeleton datasets: WH-SYMMAX [Shen et al., 2016a], SK-SMALL [Shen et al., 2016b], SK-LARGE1 [Shen et al., 2017] and SYM-PASCAL [Ke et al., 2017]. Images of these datasets are selected from other semantic segmentation datasets with human annotated object segmentation masks, and the skeleton ground-truths are extracted from segmentations. ... 1http://kaiz.xyz/sk-large |
| Dataset Splits | No | The paper specifies training hyperparameters (e.g., base learning rate, mini-batch size, momentum, maximal iteration) and data augmentation techniques but does not explicitly state the use of a validation dataset split or its size/proportion. |
| Hardware Specification | No | The paper mentions 'Limited by the memory of our GPUs' but does not provide specific details such as GPU models, CPU models, or memory specifications. |
| Software Dependencies | No | The paper states 'We implement the proposed architecture based on the openly available caffe [Jia et al., 2014] framework' but does not specify any version numbers for Caffe or other software dependencies. |
| Experiment Setup | Yes | The hyperparameters and corresponding values are: base learning rate (10 6), mini-batch size (1), momentum (0.9) and maximal iteration (40000). We decrease the learning rate every 10,000 iterations with factor 0.1. We perform the same data augmentation operations with FSDS for fair comparison. The augmentation operations are: (1) random resize images (and gt maps) to 3 scales (0.8, 1, 1.2), (2) random left-right flip images (and gt maps); and (3) random rotate images (and gt maps) to 4 angles (0, 90, 180, 270). |