Boundary Proposal Network for Two-Stage Natural Language Video Localization

A short while ago, a job of Pure Language Video clip Localization has been proposed. Provided an untrimmed movie and a pure language query, the job localizes the movie phase relevant to the query by figuring out the get started point and the finish point. A new paper proposes a two-stage finish-to-finish frame-perform termed Boundary Proposal Network. It inherits the merits of previous techniques and avoids their problems.

Picture credit history:, CC0 Public Area

Firstly, quite a few superior-quality phase proposals are produced. Then, an specific classifier matches the proposals with the sentence by predicting the matching score. The method is adaptable to movie segments of arbitrary size. The phase-amount movie element and query element are modeled jointly to increase the general performance.

The proposed network outperforms the state-of-the-art ways on three benchmark datasets. Furthermore, it is a common paradigm the place can be specific modules can be replaced by other powerful techniques.

We goal to handle the dilemma of Pure Language Video clip Localization (NLVL)-localizing the movie phase corresponding to a pure language description in a prolonged and untrimmed movie. State-of-the-art NLVL techniques are pretty much in 1-stage style, which can be generally grouped into two classes: 1) anchor-based method: it 1st pre-defines a collection of movie phase candidates (e.g., by sliding window), and then does classification for each prospect 2) anchor-free method: it right predicts the possibilities for each movie frame as a boundary or intermediate frame inside of the good phase. Even so, both of those types of 1-stage ways have inherent disadvantages: the anchor-based method is susceptible to the heuristic regulations, even further limiting the ability of handling video clips with variant size. Even though the anchor-free method fails to exploit the phase-amount interaction therefore accomplishing inferior outcomes. In this paper, we suggest a novel Boundary Proposal Network (BPNet), a common two-stage framework that gets rid of the challenges talked about previously mentioned. Specially, in the 1st stage, BPNet makes use of an anchor-free model to crank out a group of superior-quality prospect movie segments with their boundaries. In the next stage, a visible-language fusion layer is proposed to jointly model the multi-modal interaction in between the prospect and the language query, adopted by a matching score score layer that outputs the alignment score for each prospect. We appraise our BPNet on three difficult NLVL benchmarks (i.e., Charades-STA, TACoS and ActivityNet-Captions). Considerable experiments and ablative scientific tests on these datasets display that the BPNet outperforms the state-of-the-art techniques.

Analysis paper: Xiao, S., “Boundary Proposal Network for Two-Stage Pure Language Video clip Localization”, 2021. Link: muscles/2103.08109

Maria J. Danford

Next Post

ORStereo: Occlusion-Aware Recurrent Stereo Matching for 4K-Resolution Images

Wed Mar 17 , 2021
Stereo cameras can be a significant-resolution resource of 3D maps. On the other hand, helpful information processing is needed. Most existing stereo matching designs are trained on small-resolution information for the reason that significant-resolution pictures call for a ton of GPU memory. Consequently, a recent paper focuses on the generalization […]

You May Like