Ads-1k: Dataset for Ad Video Editing

Official dataset for "Multi-modal Segment Assemblage Network for Ad Video Editing with Importance-Coherence Reward" (ACCV 2022) [ACCV'22] [arXiv]

Dataset Overview

To train and evaluate the models for ad video editing task, we collect 1000+ ad videos from the advertisers to form the Ads-1k dataset. There are 942 ad videos for training and 99 for evaluation in total. However, the annotation methods of the training set and test set are somehow different. Instead of preparing the ground-truth for each data, we annotate each video with multi-labels shown in the supplementary.

Dataset statistics. $N_{seg}$ and $N_{label}$ are respectively the average number of segments and labels of each video. $D_{seg}$ and $D_{video}$ are the average duration of a segment and a video, respectively.

	$N_{seg}$	$D_{seg}$	$D_{video}$	$N_{label}$
Training Set	13.90	2.77	34.60	30.18
Test Set	18.81	1.88	34.21	35.77
Overall	14.37	2.68	35.17	30.71

Besides, the number of annotated segment pairs and the proportion are counted. The number of coherent, incoherent, and uncertain pairs are 6988, 9551, and 2971, occupying 36%, 49%, and 15%, respectively.

Note: The ads are collected from different Chinese advertisers, both large and small. Most of them are in Chinese.

Evaluation Metrics

Imp@T

$$ Imp@T=\frac{1}{|A|}\sum_{a_i\in A}imp(a_i)\cdot\mathbb{I}[c_1\cdot T\leq\tau(A)\leq c_2\cdot T]~, $$

$$ \\ imp(a_i)=\frac{1}{|L_{a_i}|}\sum_{\ell\in L_{a_i}}w_{l}\cdot \ell~, $$

where A is the set of selected segments, $L_{a_i}$ stands for the total number of labels of the selected segment, $\ell\in{1,2,3,4}$ is narrative techniques label hierarchy, and $w_{l}$ is the weight of one label. $\mathbb{I}(\cdot)$ is the indicator function. $c_1$ and $c_2$ are two constant that produced a interval based on given target duration $T$. We set $c_1=0.8$ and $c_2=1.2$, since a post-processing of $0.8\times$ slow down $1.2\times$ or fast forward can resize the result close to the target $T$ without distortion in practice. Take $T=10$ as example, the interval will be $[8,12]$, which means if the duration of result $\tau(A)=\sum_{a_i\in A}dur(a_i)\in [8,12]$ then this result is valid and gain the score.

Coh@T

The coherence score given the target duration $T$ is defined as follows:

$$ Coh@T=\frac{1}{|A|-1}{\sum_{a_i\prec a_j}coh{(a_{i},a_{j})}}\cdot\mathbb{I}[c_1\cdot T\leq\tau(A)\leq c_2\cdot T]~, $$

where the $coh{(a_{i},a_{j})}$ is the coherence small score between the text of segment $i$ and the text of segment $j$. With the annotation for coherence on test set, we can score the results produced by our models. When scoring, for each combination of consecutive two segments $i$ and $j$ in output, if it is in the coherent set, the $coh{(a_{i},a_{j})}$ will be 1. If it is in the incoherent set, the $coh{(a_{i},a_{j})}$ will be 0. Otherwise, it is in the uncertain set, the $coh{(a_{i},a_{j})}$ will be 0.5.

Imp-Coh@T

The overall score is defined as follows:

$$ ImpCoh@T=\frac{Imp@T}{|A|-1}\cdot{\sum_{a_i\prec a_j}coh{(a_{i},a_{j})}}. $$

The score reflects the ability of trade-off among importance, coherence and total duration.

Data

We have the following dataset files under the data directory:

• coh_anno_test.json: The annotation file of coherence for 99 data in test set.

• data_info.json: The information of 942 ad videos data for training.

• seg_labels_test.json: The segments with narrative techniques labels of each video in test set.

• test_info.json: The duration information of 99 ad videos data for testing

• bert_feats_test.pkl: The features of text contents (subtitles) extracted by BERT.

• swin_feats_test.pkl: The features of visual infomation (frames) extracted by Swin-Transformer (Large) from videos for test.

• vggish_feats_test.pkl: The features of audios extracted by Vggish from videos for test.

We also have the following pre-extracted segment-level features of training data, which can be downloaded from [Google Drive] or [百度网盘(提取码：8gjb)]:

• bert_feats_train.pkl: The features of text contents (subtitles) extracted by BERT from videos for training.
• swin_feats_train.pkl: The features of visual infomation (frames) extracted by Swin-Transformer (Large) from videos for training.
• vggish_feats_train.pkl: The features of audios extracted by Vggish from videos for training.
• ppl_maps.pkl: The PPL maps of training data.

Evaluation

Under the scripts directory, we include:

• eval.py: The evaluation script. Run test.py to use it.

• load_ads1k.py: The data loader for Ads-1k dataset.

• test.py: run this file to evaluate your results. Replace the nparrayyour_results in line 5 by your output:

  ...
  infer = Eval()
  your_results = ... # replace by your results
  given_times = [10,15]
  ...

Acknowledgement

We thank Southern University of Science and Technology and Tencent for support to the project. This work is supported by the National Natural Science Foundation of China under Grant No. 61972188 and 62122035.

Citation

If our work can help you, please feel free to cite it.

@InProceedings{Tang_2022_ACCV,
    author    = {Tang, Yunlong and Xu, Siting and Wang, Teng and Lin, Qin and Lu, Qinglin and Zheng, Feng},
    title     = {Multi-modal Segment Assemblage Network for Ad Video Editing with Importance-Coherence Reward},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {December},
    year      = {2022},
    pages     = {3519-3535}
}