Hi,

I would like to know if you can provide weights of your model or pretrained model ?

Respected Sir,
I know that you are using the vcdb.pickle file for the training of DML model. i also know that the pickle file is a dictionary containing video pairs and index.
i was wondering how can i generate the training files for my own dataset?

I can not load pretrained vgg model which you provided below on intermediate-cnn-feature，is there another network code?
https://drive.google.com/drive/folders/1swgdUsYwOy9pdoJzKfthAq74CqBEQDmA

Was the evaluate function in util.py created by yourself?
@gkordo

Hello, how about the way to split cc_web_video dataset to trainset and testset.
Thanks

First, thank you for your amazing work.
When running the evaluation part:
python evaluation.py --evaluation_set /data/p01/NDVR/CC_WEB_triplet/cc_web_video_features.npy --model_path /data/p01/NDVR/model/, I met error below:

Evaluation Results
==================
evaluation.py:43: RuntimeWarning: invalid value encountered in divide
sim = np.round(1 - dist[i] / dist.max(), decimals=6)
Traceback (most recent call last):
File "evaluation.py", line 78, in
positive_labels=args['positive_labels'], all_videos=False)
File "/home/p01/projects/NDVR/ndvr-dml-master/utils.py", line 157, in evaluate
precision, recall, thresholds = precision_recall_curve(y_target, y_score)
File "/home/p01/.conda/envs/pc2/lib/python2.7/site-packages/sklearn/metrics/ranking.py", line 522, in precision_recall_curve
sample_weight=sample_weight)
File "/home/p01/.conda/envs/pc2/lib/python2.7/site-packages/sklearn/metrics/ranking.py", line 416, in _binary_clf_curve
raise ValueError("Data is not binary and pos_label is not specified")
ValueError: Data is not binary and pos_label is not specified

I check the README.md and found that the order should be python evaluation.py --evaluation_set output_data/cc_vgg_features.npy --model_path model/ , so I think maybe I misunderstand the cc_vgg_features.npy, can you give me a instruction about how to build this npy file or there is something wrong at other part?

@gkordo Hi Giorgos,

I have read your paper "FIVR: Fine-grained Incident Video Retrieval" and I'm interested in the dataset of FIVR-200K which you will release ndd.iti.gr/fivr.html. Would you like to tell me when to release the FIVR-200K so I can use the dataset to do some experiments.

Thanks.

Is there any description or generate method for pickle files?

#17
It seems like it's an issue of scikit-learn and blas. Try to update/reinstall these two packages:

I am trying to update the package with the same issue as above.
By the way, blas gives an error.

pip3 install blas
requests.exceptions.HTTPError: 404 Client Error: Not Found for url: https://pypi.org/simple/blas/

Is there any other way?

Error ：
CC_WEB_VIDEO Triplet Generation

Query 0: 0%| | 0/341 [00:00<?, ?it/s]
Traceback (most recent call last):
File "triplet_generator.py", line 186, in
triplets = triplet_generator_cc(dataset, features)
File "triplet_generator.py", line 141, in triplet_generator_cc
pair_distance = euclidean(video1, video2)
File "/home/chq/anaconda3/envs/python2/lib/python2.7/site-packages/scipy/spatial/distance.py", line 602, in euclidean
return minkowski(u, v, p=2, w=w)
File "/home/chq/anaconda3/envs/python2/lib/python2.7/site-packages/scipy/spatial/distance.py", line 505, in minkowski
dist = norm(u_v, ord=p)
File "/home/chq/anaconda3/envs/python2/lib/python2.7/site-packages/scipy/linalg/misc.py", line 145, in norm
return nrm2(a)
_fblas.error: (offx>=0 && offx<len(x)) failed for 2nd keyword offx: dnrm2:offx=0

occurs when I run triplet_generator.py, do you know why? And only generates cc_web_video_features.npy
without cc_web_video_triplets.npy
thanks!

Thank you for sharing such great information on training and evaluation of your model.
i was wondering if we need to visualize what videos are near duplicates of our input test video, how this can be done using your code?

I'm trying to calculate the distance between 2 individual videos. According to the paper, after extracting the feature vectors I've got a choice between early fusing and late fusion to turn the feature vectors into a single vector before feeding it to the embedding function, but I don't see where the process even happens.

I download the feature and triplets form "https://drive.google.com/drive/folders/1swgdUsYwOy9pdoJzKfthAq74CqBEQDmA";
use vcdb_features.npy and vcdb_triplets.npy for train_dml;
the command is "python train_dml.py --train_set output_data/vcdb_features.npy --triplets output_data/vcdb_triplets.npy --model_path model/";
but it report system error:
can you help to resolve this problem ?