A new model for choosing camera views in multi-camera television footage scored 56.60% on the paper's main benchmark measure, Precision@0.5, compared with 37.16% for the prior best result. The figures come from TVMCE, which contains 88 hours of raw video contributing to 14 hours of edited footage from concerts, sports, gala shows and contests.
Published as arXiv version 1 on 26 Aug 2026, the paper asks whether separating temporal history from the evaluation of candidate views can improve recommendations. At each editing boundary, the benchmark presents 16 past shots and six candidate views, then asks the system to select the optimal camera view.
Two paths through the footage
The proposed architecture splits the job between two Transformer pathways. A dedicated temporal encoder keeps the past frames in memory, while candidate views query that memory through cross-attention, so the recommendation is made with the sequence history in view. This is the paper's central design question: whether temporal memory should be decoupled from the candidate-selection step.
For the highlighted configuration, the implementation froze a Swin Transformer V2-Tiny visual backbone, or feature extractor, and projected its features into a 768-dimensional latent space. Training used Focal Loss, a way of weighting errors during learning, for 30 epochs with AdamW at a learning rate of 1 × 10−5; the loss parameters were alpha 0.25 and gamma 2.0.
The reported scores
The highlighted SwinV2-Tiny plus Focal Loss setup reached 69.65% Precision@0.5. That result sits alongside the paper's broader comparison, in which the proposed approach scored 56.60% against 37.16% for the prior best result.
The authors also report results after adjusting the cutoff used to turn model scores into camera choices. With the macro threshold set at 0.3 after validation, final test precision was 74.66%, recall was 80.35% and F1 was 76.52%. At the reported micro threshold of 0.4, the figures were 77.52%, 74.75% and 76.11%, respectively. Recall@1 was 76.31%.
Testing quick adaptation
The paper's adaptation experiment asked how much data from a target video was needed to capture its camera-selection behavior. With no target-video fine-tuning, Precision@0.5 was 74.44% on video_0000 and 81.76% on video_0001. Using 20% of one video's training data produced approximately 84% precision.
In a cross-video test, fine-tuning on video_0001 and evaluating video_0000 produced 83.65% at the 20% mark. The result suggests that a model trained on one video may retain useful information when tested on another, but the adaptation evidence is limited to these two named high-resolution test-set videos.
The authors interpret the results as support for separating temporal history from candidate evaluation as a design choice. But the scores are tied to the TVMCE benchmark and the reported adaptation checks on video_0000 and video_0001, so they do not answer how the system would perform beyond those test settings.
The work reports support from Grant PID2022-138721NB-I00 and 2025-DI-00035 through the named Spanish and European funding bodies.
Paper data and sources
Original title: A Dual-Transformer for Multi-Camera View Recommendation
Authors: Josep Cabacas-Maso, Carles Ventura, Ismael Benito-Altamirano
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text