shyamal-b / sst Goto Github PK

Thanks for you share this excellent reseach!
I'm confused some line in 'model.py'.

input_data : ndarray
Must be three dimensional, where first dimension is the number of
input video stream(s), the second is the number of time steps, and
the third is the size of the visual encoder output for each time
step. Shape of tensor = (n_vids, L, input_size).
# --- (n_vids,L=video_length,D=500)=(1,video_length,500) ---

I ran your codes and pre_trained model on THUMOS14 , and then updated results in recall_eval.ipynb, and then plot_results.ipynb.

The curve 'sst_demo' (using sst_demo_th14_k32.hkl for predicting) is much lower than DAPs and SST in the average_recall figure. The highest average recall achieves 0.588 while 0.637 in your figure.

I am wondering if when predicting each video is the input_data size should be (1,video_length,500) according to your paper.

Thanks for you share this excellent reseach!
I'm confused some line in 'generate_sst_props.py'

def parse_args():
...
p.add_argument('-ds', '--input-dataset', help='filepath for input dataset.',
default='data/feats/example_feats.hdf5', type=str)

where can I get this file "data/feats/example_feats.hdf5". I really hope for your reply sincerely.
Best Regards!

I'm glad to see your excellent work and sorry to bother you
I just ask a stupid question

Hi, as mentioned in your paper u will release the supplementary material in this website, could u tell me where is it? Thx.

can not download the model from
https://dl.dropboxusercontent.com/s/81hlnp2g7p5gqpq/sst_demo_th14_k32.hkl

the log is ：
$wget https://dl.dropboxusercontent.com/s/81hlnp2g7p5gqpq/sst_demo_th14_k32.hkl
--2017-10-11 23:12:16-- https://dl.dropboxusercontent.com/s/81hlnp2g7p5gqpq/sst_demo_th14_k32.hkl
Resolving dl.dropboxusercontent.com... 93.46.8.89
Connecting to dl.dropboxusercontent.com|93.46.8.89|:443... failed: Connection timed out.
Retrying.

--2017-10-11 23:13:22-- (try: 2) https://dl.dropboxusercontent.com/s/81hlnp2g7p5gqpq/sst_demo_th14_k32.hkl
Connecting to dl.dropboxusercontent.com|93.46.8.89|:443... failed: Connection timed out.
Retrying.

is the link avaliabe ？

I have trained the model on Activity Net Captions and inspected its output manually and the predictions are very good. When I measured it recall with the trained model and the same model using random weights I got the same results. From my understanding, In order to measure the recall I should take top 1k proposals from each video. Is that correct or should I take the top 1k proposals from the entire dataset ?

I'd like to try train THUMOS14 video datasets as you did.
But I'm embarrassed that the video input type is not .avi or .mp4 but .hdf5 format.
I'm not familiar with .hdf5 format, so requesting you some converting tip.

Besides, I downloaded THUMOS15 video 'Test Data(untrimmed)' as your journal.
Sadly, to unzip the data, it is required password, but the THUMOS Challenge 2014 is closed a few years ago. So, if you know I request the password as well! Thanks.

Best,

dongheon Lee

Hi Shyamal, will the supplementary material be available soon? I noticed some hyper-params (e.g., k) are not specified in the full paper. Also, how about the source code? Thank you!

hello,

As you mentioned in your paper that
"sequence encoder outputs k proposals at each time step t with a confidence vector ct, where the longest proposal is of length δ · k. "

My question is if a time step have size of δ (16) frames then how the longest proposal can have size of δ.k??

kindly can you explain this? I tried to understand many times but failed :(

It seems that executing the demo proposal code requires filter.csv file which is not included in the repo. Where I can get that and what does it do?

When apply the pretrained model to ActivityNet data, due the video feature sample strategy, sometimes we will obtain video feature sequence with length 0, which will cause the Floating point exception (core dumped) error. My solution is check the length of video_feat every eval_forward and omit those videos with length 0. Hope that the author may read this issue and consider providing a new sample strategy. Also hope that this helps for those encountering the same error.

Thanks for good research!
(1) I'm wondering some part in 'generate_sst_props.py'

def parse_args():
...
p.add_argument('-ff', '--filter-file', help='filter file for dataset.',
default='th14_filter.csv', type=str)

What is the 'th14_filter.csv'?

(2) 'stride' variable is not declared in 'sst/utils.py'

Best,

dongheon Lee

Hi,

Thank you for releasing proposals for the test set videos from Thumos14.
Would it be possible to release the proposals you have obtained for the validation set as well? It would help with building localization on top of the proposals.

Thanks,
Archith