One-Shot-Learning Gesture Recognition using HOG-HOF Features
Authors: Jakub Konecny, Michal Hagara
JMLR 2014 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | All our experiments were conducted on an Intel Core i7 3610QM processor, with 2 4GB DDR3 1600 MHz memory. The running time of SM was approximately 115% of real-time (takes longer to process than to record), while MM was approximately 90% of real-time. However, none of our methods could be trivially converted to an online method, since we need to have the whole video in advance. The performance of our methods on all available data sets is presented in Table 1. |
| Researcher Affiliation | Academia | Jakub Koneˇcn y EMAIL Michal Hagara EMAIL Koreˇspondenˇcn y Matematick y Semin ar Comenius University Bratislava, Slovakia |
| Pseudocode | Yes | Algorithm 1 Trimming a video |
| Open Source Code | Yes | The code is publicly available in the MLOSS repository.1 1. The code is available at https://mloss.org/software/view/448. |
| Open Datasets | Yes | The purpose of this paper is to describe one-shot-learning gesture recognition systems developed on the Cha Learn Gesture Dataset (Cha Learn). ... Details and website: http://gesture.chalearn.org/. |
| Dataset Splits | Yes | During the challenge, development batches devel01-480 were available, with truth labels of gestures provided. Batches valid01-20 and final01-40 were provided with labels for only one example of each gesture class in each batch (training set). These batches were used for evaluation purposes. |
| Hardware Specification | Yes | All our experiments were conducted on an Intel Core i7 3610QM processor, with 2 4GB DDR3 1600 MHz memory. |
| Software Dependencies | No | The paper mentions "efficient implementation from Piotr s toolbox (Doll ar), function hog(image, 40, 16)" and "MATLAB code for creating the matrix A which captures these properties is in Appendix B." However, specific version numbers for Piotr's toolbox or MATLAB are not provided, which is required for a 'Yes' answer. |
| Experiment Setup | Yes | We used a simple [ 1, 0, 1] gradient filter, applied in both directions and discretized the gradient orientations into 16 orientation bins between 0 and 180 . We had cells of size 40 40 pixels and blocks of size 80 80 pixels, each containing 4 cells. ... Let 0 m < 1 be a normalization factor. ... (in our methods set to 0.5). ... gap 3 max Diff 15 threshold 0.1 min Trim 5 for i = 1 n do ... When being in a particular node n at time t, moving to time t + 1 we can either stay in the same node (slower performance), move to node n + 1 (the same speed of performance), or move to node n+2 (faster performance). |