A computer-vision model that selects external video views reported higher benchmark accuracy when estimating proficiency from synchronized first-person and external footage. Compared with SkillFormer, the authors report a 5.5-percentage-point absolute difference on Ego-Exo4D and a 7.1-percentage-point absolute difference on EgoExo-Fitness. On Ego-Exo4D, reported gains across three pretrained video backbones were 12.7% for K400, 3.1% for K600 and 9.1% for HowTo100M.
The task is ordinal classification: the system places a video in an ordered proficiency level rather than producing a continuous score. The benchmarks pair an egocentric view with exocentric views for fusion.
Ego-Exo4D covers nine activity categories and 1,087 hours of footage from one egocentric and four exocentric views. It contains 2,987 proficiency scores across four discrete levels. EgoExo-Fitness contains 1,276 cross-view videos, about 32 hours of footage, segmented into 6,131 single actions rated on a five-level scale.
A selective approach to multiple views
The proposed design pairs weakly supervised AdaMVS token and view selection with VIB-GB. VIB-GB combines Gradient Blending with a Variational Information Bottleneck, an information-compression step, to address redundancy and overfitting in multiview features.
The selection process is adaptive rather than fixed. During training, AdaMVS uses Gumbel-Softmax differentiable weighting so the model can learn weights for candidate views; at inference, it uses Soft Fusion to retain complementary cues while filtering noise.
VIB-GB combines information compression with dynamic OGR-based weighting, linking regularisation of the learned representation to regularisation based on training dynamics. In the authors’ description, AdaMVS addresses redundancy in the input views, while VIB-GB addresses overfitting in the learned features.
The reported results
In the comparison with overfitting baselines, the full configuration with VIB-GB is listed at accuracy values of 53.0 for EgoExo4D and 46.4 for EgoExo-Fitness. No uncertainty estimate is shown alongside those values.
A smaller configuration, AdaMVS-Small, is reported at 0.26 GFLOPs and 2.19 million parameters, with similar accuracy. The paper describes this as a reduction of more than 11 times in both computation and model size.
In the component ablation, reported accuracy was 45.5% for the baseline, 48.0% with AdaMVS, 48.4% after CORN, 48.6% with GB and 51.0% with GB+VIB. VIB-GB alone on the naive baseline was reported at 48.7%, a 3.2-point increase, but 2.3 points below the full model.
The training curves provide a descriptive overfitting signal. The baseline was reported to overfit around the 10th epoch and had 48.7% accuracy; with VIB-GB, the reported accuracy was 53.0%. With VIB-GB, OGR rose slightly before the 10th epoch, then decreased and stabilised after around the 20th epoch.
Two views were the reported high point
AdaMVS had its highest reported accuracy when it selected K = 2 exocentric views. The paper says two dominant views account for approximately 82.8% of all videos and describes additional views as mainly redundant or noisy.
A supplementary redundancy analysis reports a matching descriptive pattern: views with low Rv values consistently received higher AdaMVS importance weights. The paper does not report an inferential test of that correspondence.
A supplementary control is harder to interpret. The prose says that repeating the same external view four times, in the Same Exo×4 condition, reduced performance. But the reported values show Summation rising from 45.7 to 46.6 and AdaMVS from 45.9 to 46.6; AdaMVS with all exocentric views is listed at 49.2.
Why the findings remain provisional
The authors identify subjective proficiency annotations and limited dataset scale as limitations. The reported benchmark differences also come without confidence intervals or formal significance tests, so the analysis does not quantify the uncertainty around those comparisons.
The study evaluates computational models on the two named video benchmarks rather than human participants. It does not show improved human proficiency, training outcomes or real-world assessment quality, and it does not establish causal effects in real-world assessment.
The manuscript is an arXiv version-one computer-vision preprint dated 26 August 2026.
Paper data and sources
Original title: Moving Beyond More Views: Redundancy-Aware Ego-Exo Fusion for Proficiency Estimation
Authors: Xu Dong, Wanqing Li, Anthony Adeyemi-Ejeye, Andrew Gilbert
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text