Searcharxiv⌕ Search

arXiv subjects

Cristian Gavrus

Publications and source records attributed to Cristian Gavrus.

8 recordsLinked to original sources
arXiv2025

Robust Context-Aware Object Recognition

In visual recognition, both the object of interest (referred to as foreground, FG, for simplicity) and its surrounding context (background, BG) play an important role. However, standard supervised learning often leads to unintended over-reliance on the BG, known as shortcut learning of spurious correlations, limiting model robustness in real-world deployment settings. In the literature, the problem is mainly addressed by suppressing the BG, sacrificing context information for improved generalization. We propose RCOR -- Robust Context-Aware Object Recognition -- the first approach that jointly achieves robustness and context-awareness without compromising either. RCOR treats localization as an integral part of recognition to decouple object-centric and context-aware modelling, followed by a robust, non-parametric fusion. It improves the performance of both supervised models and VLM on datasets with both in-domain and out-of-domain BG, even without fine-tuning. The results confirm that localization before recognition is now possible even in complex scenes as in ImageNet-1k.

cs.CV↗
arXiv2025

Bringing the Context Back into Object Recognition, Robustly

In object recognition, both the subject of interest (referred to as foreground, FG, for simplicity) and its surrounding context (background, BG) may play an important role. However, standard supervised learning often leads to unintended over-reliance on the BG, limiting model robustness in real-world deployment settings. The problem is mainly addressed by suppressing the BG, sacrificing context information for improved generalization. We propose "Localize to Recognize Robustly" (L2R2), a novel recognition approach which exploits the benefits of context-aware classification while maintaining robustness to distribution shifts. L2R2 leverages advances in zero-shot detection to localize the FG before recognition. It improves the performance of both standard recognition with supervised training, as well as multimodal zero-shot recognition with VLMs, while being robust to long-tail BGs and distribution shifts. The results confirm localization before recognition is possible for a wide range of datasets and they highlight the limits of object detection on others

cs.CV↗
arXiv2022

The gauge-invariant I-method for Yang-Mills

We prove global well-posedness of the $ 3d $ Yang-Mills equation in the temporal gauge in $ H^σ $ for $ σ> \frac{5}{6} $. Unlike related equations, Yang-Mills is not directly amenable to the method of almost conservation laws (I-method) in its Fourier and global version. We propose a modified energy which: 1) Is gauge-invariant and easy to localize 2) Provides local gauges which give control of local Sobolev norms (through an Uhlenbeck-type lemma for fractional regularities) 3) Is slightly smoother in time compared to the classical I-method energy for related systems. The spatial smoothing is realized via the Yang-Mills heat flow instead of the multiplier $I$. Due to the temporal condition and its finite speed of propagation, the local gauge selection is compatible with recent initial data extension results. Therefore, smoothened energy differences can be partitioned into local pieces whose (appropriately extended) bounds can be square summed. After revealing the null structure within the trilinear integrals, these can be estimated using known methods. In an appendix we show how an invariant modified energy for Maxwell-Klein-Gordon can extend previous results to regularities $ σ> \frac{5}{6} $.

math.AP↗
arXiv2019

Wave maps on (1+2)-dimensional curved spacetimes

In this article we initiate the study of 1+ 2 dimensional wave maps on a curved spacetime in the low regularity setting. Our main result asserts that in this context the wave maps equation is locally well-posed at almost critical regularity. As a key part of the proof of this result, we generalize the classical optimal bilinear L^2 estimates for the wave equation to variable coefficients, by means of wave packet decompositions and characteristic energy estimates. This allows us to iterate in a curved X^{s,b} space.

math.AP↗
arXiv2017

Global well-posedness for the massive Maxwell-Klein-Gordon equation with small critical Sobolev data

In this paper we prove global well-posedness and modified scattering for the massive Maxwell-Klein-Gordon equation in the Coulomb gauge on $\mathbb{R}^{1+d}$ $(d \geq 4)$ for data with small critical Sobolev norm. This extends to the general case $ m^2 > 0 $ the results of Krieger-Sterbenz-Tataru ($d=4,5 $) and Rodnianski-Tao ($ d \geq 6 $), who considered the case $ m=0$. We proceed by generalizing the global parametrix construction for the covariant wave operator and the functional framework from the massless case to the Klein-Gordon setting. The equation exhibits a trilinear cancelation structure identified by Machedon-Sterbenz. To treat it one needs sharp $ L^2 $ null form bounds, which we prove by estimating renormalized solutions in null frames spaces similar to the ones considered by Bejenaru-Herr. To overcome logarithmic divergences we rely on an embedding property of $ \Box^{-1} $ in conjunction with endpoint Strichartz estimates in Lorentz spaces.

math.AP↗
arXiv2016

Global well-posedness of high dimensional Maxwell-Dirac for small critical data

In this paper, we prove global well-posedness of the massless Maxwell-Dirac equation in Coulomb gauge on $\mathbb{R}^{1+d}$ $(d \geq 4)$ for data with small scale-critical Sobolev norm, as well as modified scattering of the solutions. Main components of our proof are A) uncovering null structure of Maxwell-Dirac in the Coulomb gauge, and B) proving solvability of the underlying covariant Dirac equation. A key step for achieving both is to exploit (and justify) a deep analogy between Maxwell-Dirac and Maxwell-Klein-Gordon (for which an analogous result was proved earlier by Krieger-Sterbenz-Tataru), which says that the most difficult part of Maxwell-Dirac takes essentially the same form as Maxwell-Klein-Gordon.

math.AP↗