VISIONE Feature Repository for VBS: Multi-Modal Features and Detected Objects from VBSLHE Dataset
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
This repository contains a diverse set of features extracted from the VBSLHE dataset (laparoscopic gynecology) . These features will be utilized in the VISIONE system [Amato et al. 2023, Amato et al. 2022] in the next editions of the Video Browser Showdown (VBS) competition (https://www.videobrowsershowdown.org/). We used a snapshot of the dataset provided by the Medical University of Vienna and Toronto that can be downloaded using the instructions provided at https://download-dbis.dmi.unibas.ch/mvk/. It comprises 75 video files. We divided each video into video shots with a maximum duration of 5 seconds. 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 (v2) 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 VBSLHE 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. 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 VBSLHE 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 (see links above). 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/CV
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.002 | 0.000 |
| Scholarly communication | 0.004 | 0.001 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it