VISIONE Feature Repository for VBS: Multi-Modal Features and Detected Objects from V3C1+V3C2 Dataset

This repository contains a diverse set of features extracted from the V3C1+V3C2 dataset, sourced from the Vimeo Creative Commons Collection. These features were utilized in the VISIONE system [Amato et al. 2023, Amato et al. 2022] during the latest editions of the Video Browser Showdown (VBS) competition (https://www.videobrowsershowdown.org/).

The original V3C1+V3C2 dataset, provided by NIST, can be downloaded using the instructions provided at https://videobrowsershowdown.org/about-vbs/existing-data-and-tools/.

It comprises 7,235 video files, amounting for 2,300h of video content and encompassing 2,508,113 predefined video segments.

We subdivided the predefined video segments longer than 10 seconds into multiple segments, with each segment spanning no longer than 16 seconds. As a result, we obtained a total of 2,648,219 segments. For each segment, we extracted one frame, specifically the middle one, and computed several features, which are described in detail below.

This repository is released under a Creative Commons Attribution license. If you use it in any form for your work, please cite the following paper:

@inproceedings{amato2023visione, 
title={VISIONE at Video Browser Showdown 2023}, 
author={Amato, Giuseppe and Bolettieri, Paolo and Carrara, Fabio and Falchi, Fabrizio and Gennaro, Claudio and Messina, Nicola and Vadicamo, Lucia and Vairo, Claudio}, 
booktitle={International Conference on Multimedia Modeling}, 
pages={615--621}, 
year={2023}, 
organization={Springer} 
}

This repository comprises the following files:

• msb.tar.gz  contains tab-separated files (.tsv) for each video. Each tsv file reports, for each video segment, the timestamp and frame number marking the start/end of the video segment, along with the timestamp of the extracted middle frame and the associated identifier ("id_visione"). 
• extract-keyframes-from-msb.tar.gz contains a Python script designed to extract the middle frame of each video segment from the MSB files. To run the script successfully, please ensure that you have the original V3C videos available.
• features-aladin.tar.gz^† contains ALADIN [Messina N. et al. 2022] features extracted for all the segment's middle frames. 
• features-clip-laion.tar.gz^† contains CLIP ViT-H/14 - LAION-2B [Schuhmann et al. 2022] features extracted for all the segment's middle frames.
• features-clip-openai.tar.gz^† contains CLIP ViT-L/14 [Radford et al. 2021] features extracted for all the segment's middle frames. 
• features-clip2video.tar.gz^† contains CLIP2Video [Fang H. et al. 2021] extracted for all the video segments.  In particular 1) we concatenate consecutive short segments so to create segments at least 3 seconds long; 2) we downsample the obtained segments to 2.5 fps; 3) we feed the network with the first min(36, n) frames, where n is the number of frames of the segment. Notice that the minimum processed length consists of 7 frames, given that the segment is no shorter than 3s. 
• objects-frcnn-oiv4.tar.gz^* contains the objects detected using  Faster R-CNN+Inception ResNet (trained on the Open Images V4 [Kuznetsova et al. 2020]). 
• objects-mrcnn-lvis.tar.gz^* contains the objects detected using Mask R-CNN [He et al. 2017] (trained on LVIS).
• objects-vfnet64-coco.tar.gz^* contains the objects detected using VfNet [Zhang et al. 2021] (trained on COCO dataset).

*Please be sure to use the v2 version of this repository, since v1 feature files may contain inconsistencies that have now been corrected

*Note on the object annotations: Within an object archive, there is a jsonl file for each video, where each row contains a record of a video segment (the "_id" corresponds to the "id_visione" used in the msb.tar.gz) . Additionally, there are three arrays representing the objects detected, the corresponding scores, and the bounding boxes. The format of these arrays is as follows:

• "object_class_names": vector with the class name of each detected object.
• "object_scores": scores corresponding to each detected object.
• "object_boxes_yxyx": bounding boxes of the detected objects in the format (ymin, xmin, ymax, xmax).

^†Note on the cross-modal features: The extracted multi-modal features (ALADIN, CLIPs, CLIP2Video) enable internal searches within the V3C1+V3C2 dataset using the query-by-image approach (features can be compared with the dot product). However, to perform searches based on free text, the text needs to be transformed into the joint embedding space according to the specific network being used. Please be aware that the service for transforming text into features is not provided within this repository and should be developed independently using the original feature repositories linked above.

We have plans to release the code in the future, allowing the reproduction of the VISIONE system, including the instantiation of all the services to transform text into cross-modal features. However, this work is still in progress, and the code is not currently available.

References:

[Amato et al. 2023] Amato, G.et al., 2023, January. VISIONE at Video Browser Showdown 2023. In International Conference on Multimedia Modeling (pp. 615-621). Cham: Springer International Publishing.

[Amato et al. 2022] Amato, G. et al. (2022). VISIONE at Video Browser Showdown 2022. In: , et al. MultiMedia Modeling. MMM 2022. Lecture Notes in Computer Science, vol 13142. Springer, Cham. 

[Fang H. et al. 2021] Fang H. et al.,  2021. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097.

[He et al. 2017] He, K., Gkioxari, G., Dollár, P. and Girshick, R., 2017. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 2961-2969).

[Kuznetsova et al. 2020] Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A. and Duerig, T., 2020. The open images dataset v4. International Journal of Computer Vision, 128(7), pp.1956-1981.

[Lin et al. 2014] Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P. and Zitnick, C.L., 2014, September. Microsoft coco: Common objects in context. In European conference on computer vision (pp. 740-755). Springer, Cham.

[Messina et al. 2022] Messina N. et al., 2022, September. Aladin: distilling fine-grained alignment scores for efficient image-text matching and retrieval. In Proceedings of the 19th International Conference on Content-based Multimedia Indexing (pp. 64-70).

[Radford et al. 2021] Radford A. et al., 2021, July. Learning transferable visual models from natural language supervision. In International conference on machine learning (pp. 8748-8763). PMLR.

[Schuhmann et al. 2022] Schuhmann C. et al., 2022. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35, pp.25278-25294.

[Zhang et al. 2021] Zhang, H., Wang, Y., Dayoub, F. and Sunderhauf, N., 2021. Varifocalnet: An iou-aware dense object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 8514-8523).