escorciav / deep-action-proposals Goto Github PK

I don't understand how much proposals are generated in all in one video.During training, only one stream is processed in one video,so the overall number of proposals is K.Is that true?

compute_priors function on data_generation module is a little bit convoluted. The matching of priors to segments should be on other function because we will end-up duplicating code if all the segments of the dataset are not used to compute priors. That is the case of ActivityNet where there is an explicit validation set.

Similar refactoring should be done on tool create_dataset.py

We need a module to interact with activity-net dataset. I would recommend to make something similar to thumos14_helper.Thumos14 that generate the CSV-files required to interact with the annotations.

This loop assumes that the DataFrame is contiguous i.e. index from 0 to df.shape[0].

Some videos of Thumos14 are incorrectly annotated e.g. video_validation_0000364 and video_validation_0000856. In order to avoid unpleasant surprises, it's useful to drop rows of the data-frame where t-init or t-end < video-duration

idx_drop = ((df['video-duration'] < df['t-init']) | (df['video-duration'] < df['t-end'])).nonzero()[0]
df.drop(idx_drop, inplace=True)

Python program similar to bash-source used in c3d feature extraction example.

• Create output dir if it doesn' exist.
• Modify prototxt

The function ReadImageSequenceToVolumeDatum on C3D is hard-coded to read images with format %s/%06d.jpg therefore we should dump frames of videos accordingly.

this line should be f_init + t_size- 1 because the end-frame should be contained on the segment according to segment utilities.

this line imposes a constraint on the initial-frame.

Note: indexing of features with this convention will require an increment on end-frame.
@cabaf

Handle multiple process on the same GPU may hurt the run-time performance of each process. In that order of ideas, it is useful to dispatch several tasks to the same GPU in serial mode.

Bug: confidence matrix is not dumped by create_dataset therefore it is not possible to train a model

hi,
i am geting "no such a file or directory" message after runing tools programm batch_frame_extraction.py

I wrote this function replicating the idea of the cvpr-paper. It would be great to refine it in the following aspects:

• Apply intersection-threshold less than 1 just on annotations with length similar or greater that t_size.
• Segments with high coverage ratio should be more likely to be selected i.e. sampling based on cov_ratio_per_segment.
• define appropriate values for RATIO_INTERVALS
• main bug is here. index i is iterating over videos while score.shape[0] corresponds to the number of segments for all the videos. My fault 😞

Note: if t_size is closer to l_size is possible that not enough segments are generated to satisfy a uniform distribution on RATIO_INTERVALS. We should decide how to overcome it.

We are interested on retrieve c3d features associated with a segment of interest. Given a video segment, a feature stride, a folder with features of the video.

• compute list with frames of interest
• read features on previous list
• stack features
• apply a pooling operator, optional
• save stack, optional

Apparently there is a bug in lasagne or theano that produces as error if the input of the LSTM is an array of 3-dimesions and sizes of 1 in the first dimension array.shape = (1, X, X).