search

Finspire13 / CMCS-Temporal-Action-Localization

Completeness Modeling and Context Separation for Weakly Supervised Temporal Action Localization (CVPR2019)
MIT License
150 stars 17 forks source link

thumos14, extract i3d features #6

Closed Zhuysheng closed 5 years ago

Zhuysheng commented 5 years ago

When I try to extract features on thumos14 dataset by the default setting, the length of feature is less than yours, how do you set the chunk size and frequency

Finspire13 commented 5 years ago

The frequency is every 4 or 5 frames (so the snippets are overlapping). Please refer to the config

huge123 commented 3 years ago

@Finspire13 In this case that the snippets are overlapping, why the paper says " I3D is pre-trained on Kinetics [7] and takes non-overlapping 16-frame chunks as input for both two streams" in section 4.2?

Finspire13 commented 3 years ago

@Finspire13 In this case that the snippets are overlapping, why the paper says " I3D is pre-trained on Kinetics [7] and takes non-overlapping 16-frame chunks as input for both two streams" in section 4.2?

@huge123 Hi, if I remember correctly, the feature snippets are overlapping at their extraction (please refer to 'base_sample_rate' in the configs), and the feature sent into the model is non-overlapping (please refer to 'sample_rate' in the configs). Such difference make it possible to augment the data temporally when training (please refer this code.