Action Localization with Tubelets from Motion

By Mihir Jain, Jan van Gemert, Herv'e J'egou, Patrick Bouthemy, Cees G. M. Snoek. 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014

304 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/377ad65969b98823dc5f28815d8a01b74fc1b79a

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Weakly supervised learning of actions from transcripts (2016, background): “Some approach the problem as a localization task in the spatiotemporal domain [12, 13, 14] and cluster trajectories or apply dynamic programming to enable efficient search for actions in the input space.”
Spot On: Action Localization from Pointly-Supervised Proposals (2016, methodology): “Recent work using unsupervised action proposals based on supervoxels [2,5,19] or on trajectory clustering [4,20,21], have shown good results for action localization.”
MAN: Moment Alignment Network for Natural Language Moment Retrieval via Iterative Graph Adjustment (2018, methodology): “Earlier works on activity detection mainly used temporal sliding windows as candidates and trained activity classifier on hand-crafted features [35, 13, 23, 33, 46].”
Human Action Recognition and Prediction: A Survey (2018, background): “[71] proposed to generate a set of bounding boxes from the video which are called tubelets for action localization.”
SCSampler: Sampling Salient Clips From Video for Efficient Action Recognition (2019, background): “More closely related to our task is the problem of action localization [22, 39, 38, 54, 61], where the objective is to localize the temporal start and end of each action within a given untrimmed video and to recognize the action class.”
Unsupervised Object Discovery and Tracking in Video Collections (2015, methodology): “Finally, spatio-temporal proposals of [16, 24] and action localization [39, 42] are relevant to our work as they also return spatio-temporal tubes as output.”
Towards Weakly-Supervised Action Localization (2016, unknown): “[23] construct action tubes by hierarchically merging supervoxels and rely on dense trajectory features for tube classiﬁcation.”
Learning to Track for Spatio-Temporal Action Localization (2015, methodology): “g the same protocol, we obtain a performance of 91.9% for UCF-Sports and 57.99% for J-HMDB using state-of-the-art improved dense trajectories [44] with Fisher ours Gkioxari and Malik [11] Jain et al. [15] Wang et al. [45] Tian et al. [39] Lan et al . [27] IOU threshold AUC 0.1 0 0.1 0.2 0.2 0.3 0.3 ”
Object Detection from Video Tubelets with Convolutional Neural Networks (2016, methodology): “[13] adopted the Selective Search strategy on super-voxels to generate tubulet proposals and proposed new features to differentiate human actions from background movements.”
Action Sets: Weakly Supervised Action Segmentation Without Ordering Constraints (2017, methodology): “When processing untrimmed videos, actions can either be localized in the temporal domain only [2, 24, 39, 29, 25, 15, 17, 26], or in the spatio-temporal domain [33, 7, 20, 38].”
SST: Single-Stream Temporal Action Proposals (2017, unknown): “[4, 12, 16, 34, 37, 39].”
Efficient Action Detection in Untrimmed Videos via Multi-task Learning (2016, unknown): “Presently, the dominant approach to action detection [10, 12, 43, 28] is a sequential pipeline usually consisting of three steps, either using IDT features plus shallow classifiers or through deep neural networks.”
Tube Convolutional Neural Network (T-CNN) for Action Detection in Videos (2017, unknown): “[6] and Soomro et al .”
Human Action Localization with Sparse Spatial Supervision (2017, methodology): “[29] build action localization candidates by hierarchically merging supervoxels and use dense trajectory features for tube classification.”
Action Tubelet Detector for Spatio-Temporal Action Localization (2017, background): “Hundreds of action proposals are extracted per video given low-level cues, such as super-voxels [11, 22] or dense trajectories [3, 7, 20].”
Encouraging LSTMs to Anticipate Actions Very Early (2017, background): “This is in contrast to existing methods that typically extract either global representations for the entire image [6, 41, 42, 7] or video sequence [35, 14], thus not focusing on the action itself, or localize the feature extraction process to the action itself via dense trajectories [39, 38, 9], opt”
What do 15,000 object categories tell us about classifying and localizing actions? (2015, unknown): “Instead of localizing cuboids [2], recently action location is more precisely defined as a sequence of bounding boxes [15, 22].”
Finding action tubes (2014, methodology): “Our approach outperforms all other approaches ([15, 42, 38, 25]) on UCF sports, with the biggest gain observed for high overlap thresholds.”