models

This folder contains DAPs models.

T512K64 THUMOS14

Model used in the camera ready of our ECCV-work. It was trained on 80% of THUMOS14 validation set with temporal annotations.

Details

format: This file was save numpy.savez using the version reported in our environment YAML file.

order: the order of the arrays comes from the function get_all_param_values applied on our sequence encoder with default values.

anchors used to train the model T512K64 THUMOS14.

Details

format: This file is an HDF5 file with a unique dataset called anchors.

Model used in an early-stage version of our work. It was trained on 80% of THUMOS14 validation set with temporal annotations.

Details

format: This file was save numpy.savez using the version reported in our environment YAML file.

anchors used to train the model T256K16 THUMOS14.

format: This file is an HDF5 file with a unique dataset called anchors.

Results of PCA analysis of C3D (FC7) representation of videos in THUMOS14.

Details

format: This file is an HDF5 file with three datasets S, U, x_mean.

Usage

Given a video-clip with a visual representation x, you can reduce its dimensionality up to 500 dimension by doing:

import numpy as np

# x = define-your-feature-vector-here
num_red_dim = 500
x_red = np.dot(x - x_mean, U[:, :num_red_dim])

Note for curious user: you can easily plug this operation as a lasagne layer 😉