arXiv subjects

Ankit Singh

Publications and source records attributed to Ankit Singh.

At least 19 recordsLinked to original sources

arXiv2026

Nonlinear visco-elasto-plastic model for the stress-strain and stress-relaxation in metallic glass

We develop a nonlinear visco-elasto-plastic (VEP) model for the stress-strain response of metallic glasses by connecting the deformation dynamics with the material relaxation time obtained from independent stress-relaxation measurements. The material is modeled as a Maxwell-type element, with a nonlinear elastic spring and a stretched-exponential dashpot whose behavior is determined from the separate stress-relaxation data. The model captures the effects of temperature and strain rate on the initial elastic response, stress overshoot, post-yield softening, and steady flow. We independently determine the reference material-time function $X_0(\varepsilon)$ from the stress-relaxation response and the nonlinear material-time variable $X(\varepsilon)$ from the elastic-viscoplastic strain decomposition based on the stress-strain response. We introduce an Eyring-type function or material-time ratio $ϕ(\varepsilon)=X(\varepsilon)/X_0(\varepsilon)$ which quantifies the change of relaxation dynamics during nonlinear deformation. Our model provides a simple connection between structural relaxation and yielding in metallic glasses.

cond-mat.soft↗

arXiv2026

Channelled versus Isotropic Accretion of Galaxy Groups onto Clusters: Kinematics, Spin Alignment, and Assembly Bias

Galaxy clusters grow by accreting groups from the cosmic web, but whether the route of infall (along a cluster-connected filament or from the open field) imprints a group's kinematics, spin, and orbit remains unmeasured at scale. We classify $212{,}704$ group-cluster infall pairs across $27{,}029$ clusters in the $(1\,h^{-1}\,\mathrm{Gpc})^3$ MDPL2 simulation into primary (cluster-connected), secondary (tributary), primary-outskirt, and isotropic channels using DisPerSE filaments. (i) In both channelled classes, group spin axes lie mainly parallel to the filament below a `spin-flip' mass and perpendicular above it, $\log_{10} M_\mathrm{flip}\;[h^{-1}\,M_{\odot}] = 13.07$ and $13.12$, a $0.05$-dex offset consistent with zero, lying $1.3-1.4$ dex above the field, while isotropic groups show no flip; $M_\mathrm{flip}$ stays radially flat out to $7\,R_{200}$. (ii) Within the primary class $M_\mathrm{flip}$ rises with the linear mass density of the host filament, from $12.96$ (thin tendrils) to $13.44$ (dense bridges); matching group mass, cluster mass and distance keeps the rise positive but leaves its size uncertain ($0.07-0.58$ dex). (iii) Isotropic groups fall below random alignment at $2\,R_{200}$ ($41$ per cent within $60^\circ$, a $4.3σ$ deficit). Primary groups show $85$ per cent velocity-filament alignment at $2\,R_{200}$, and isotropic groups infall $150-180\,\mathrm{km\,s^{-1}}$ faster, and more radially, out to $5\,R_{200}$. At matched distance isotropic groups are therefore predicted to reach orbital pericentres up to $10$ per cent smaller than primary groups, whose smaller pooled pericentres reflect only their proximity to the cluster. The topological route of accretion, and within it the host filament's linear density, shapes how groups enter clusters, testable with wide-area spectroscopic surveys.

astro-ph.GA↗

arXiv2026

Iterated Lavrentiev regularization under random noise: variational source conditions and rates in expectation

The convergence theory of iterated Lavrentiev regularization for nonlinear monotone ill-posed equations is deterministic: the data error is bounded by a known level that the discrepancy principle uses explicitly. Under Gaussian noise no such bound exists. For a discretized two-dimensional nonlinear inverse source problem for the Poisson equation, we analyse convergence in a finite-energy Gaussian noise model, keeping the linearized and nonlinear cases apart. For the linearized iteration the noisy and noise-free iterates differ by one resolvent applied to the noise; this gives an exact singular-value variance formula, a priori rates in expectation, and the main result: for a family of discretizations whose distance functions are bounded uniformly in the mesh, the rate is governed by that decay, which relaxes the classical range condition. The rate degrades gracefully, down to logarithmic, as the underlying continuum error loses smoothness, and a balancing rule attains it without knowing the smoothness. For the full nonlinear iteration under the order-one source condition, a truncated discrepancy principle fed only the noise level attains $σ_E^{1/2}$ in expectation, in the energy noise level $σ_E\asympσ/h$, the rare event of large noise being controlled by concentration. The rough nonlinear regime and the general-exponent nonlinear rate are left open and supported numerically for kink and jump errors. We separate the deterministic worst-case, Gaussian, and white-noise settings, prove a worst-case lower bound, and describe where the method is statistically inefficient relative to spectral cut-off. Heterogeneous media and a cost comparison with an adjoint-based scheme are given as supplementary material.

math.NA↗

arXiv2026

Retail Product Search: A Practical Approach at Target

Search is one of the most important features in e-commerce, directly driving customer engagement and business growth. A good product search system must show both relevant and desirable results. However, retail search presents unique challenges. User intent can range from exact matches to open-ended discovery. Search systems must also balance multiple goals, such as relevance, revenue, and profit, while keeping response times low. Traditional keyword-based methods often fall short in handling natural language or semantic queries. Vector search helps alleviate these issues, but it can miss key intent signals or return low-precision results. In this paper, we present the design of a hybrid search system at Target that combines lexical and vector search. We describe our approach to data processing, embedding training, precision control for the final result set, multi-channel result fusion (where we compared fusion strategies and adopted weighted interleaving), and the performance optimizations used to maintain low latency for production deployment. Our method improves offline evaluation metrics, and in online A/B testing it raised click-through rate by 0.97%, order conversion by 0.98%, and demand per visitor by 1.10% over lexical-only search, while roughly halving zero-result searches. The resulting system is deployed at scale and serves millions of guests daily.

cs.IR↗

arXiv2026

The influence of galaxy mergers, black-hole growth, and gas processes on the evolution of the stellar mass-gas metallicity relation of galaxies in different cosmic environments

We study the impact of supermassive black hole (SMBH) growth, $\langle \dot{M}_\mathrm{SMBH}\rangle$, major and minor galaxy mergers, and gas processes, on the average gas metallicity of galaxies within the Horizon Run 5 simulation, with the aim to uncover which of these processes drive the scatter in the gas metallicity-stellar mass relation (MZR) at different redshifts in nodes, filaments and voids. At $z=5$, minor mergers produce the largest differential in $\log[Z_g/Z_\odot]$ where all environments display a maximum $0.22$ dex increase in the average $\log[Z_g/Z_\odot]$ compared to non-merging galaxies. The node population also displays a consistent $0.1$ dex negative offset in $\mathrm{d} \log[Z_g/Z_{\odot}]$, the residual $Z_g$ from the total MZR of all galaxies, across all redshifts, whilst filament and void galaxies show a smaller offset. Major mergers show little influence on these same properties. This suggests minor mergers regulate metallicity and contribute to galaxy mass growth concurrently, accelerating chemical evolution post merger. Between $z=1-3$, a high $\langle \dot{M}_\mathrm{SMBH}\rangle$ leads to a larger negative offset in $\mathrm{d} \log[Z_g/Z_{\odot}]$ for all environments. Here, node galaxies show the largest negative offset of approximately $0.25$ dex, suggesting that AGN-driven gas removal may contribute to the MZR scatter at intermediate times. Finally, galaxies with low $M_{gas}/{M_{tot}}$ show increased $\mathrm{d} \log[Z_g/Z_{\odot}]$ across all redshifts and environments, again a $0.25$ dex maximum for node galaxies. These galaxies also spike in $\mathrm{d} \log[Z_g/Z_{\odot}]$ at late times, below $z=1$. At this time, galaxies in the nodes show negative $\langle \dot{M}_\mathrm{gas} \rangle$ whilst also showing the largest $\mathrm{d} \log[Z_g/Z_{\odot}]$ values we observe of $0.2$ dex.

astro-ph.GA↗

arXiv2026

Intracluster Light as a Probe for Dark Matter: Exploring Self-interacting Dark Matter and Cold Dark Matter with C-EAGLE Sims

We assess whether intracluster light (ICL) can serve as an observational discriminator of dark matter physics. The self-interacting dark matter (SIDM) model has gained increasing attention as a possible resolution to small-scale discrepancies between collisionless cold dark matter (CDM) simulations and observations, predicting distinct tidal interaction histories within galaxy clusters. We analyze Cluster-EAGLE zoom-in galaxy clusters re-simulated from identical initial conditions in both CDM and SIDM frameworks. The morphological similarity between dark matter and multiple baryonic tracers -- gas, all stars, galaxies, and the combined brightest cluster galaxy plus ICL (BCG+ICL) -- is quantified using the Weighted Overlap Coefficient, a contour-overlap statistic. We find that dark matter is traced most accurately by BCG+ICL, followed by gas, all stars, and galaxies. The BCG+ICL component remains a robust tracer even at high redshift, while gas initially traces dark matter poorly but improves over time, eventually approaching the performance of BCG+ICL. Notably, in the SIDM case the gas distribution more closely resembles dark matter than in CDM. This reflects the underlying physics: in CDM, collisionless dark matter behaves similarly to the collisionless BCG+ICL, whereas in SIDM, self-interactions introduce an effective collisionality, making dark matter evolve more like the gas component. We also find that dwarf and satellite galaxies are more sensitive to the underlying dark matter model, despite their poorer overall tracing performance. Our results demonstrate the potential of ICL as a novel observational probe of dark matter physics and provide a first step toward using diffuse cluster light to constrain the nature of dark matter.

astro-ph.CO↗

arXiv2026

Glassy dynamics, crossover temperature and density scaling in fragile glass-formers

We investigate the slowing down of dynamics in a glass-forming mixture interacting via an inverse-power-law (IPL) potential using a combination of theory and large-scale molecular dynamics simulations. We measure the static pair-correlation function, configurational entropy, inherent-structure energy, and structural relaxation time. We employ a theoretical framework to calculate the structural relaxation time $τ_α$, which is found to be in very good agreement with the simulation results. The theory identifies a local structural order which defines the cooperativity of the relaxation and brings forth a fluctuation induced parameter $ψ( T )$ and a crossover temperature $T_a$ that characterize the density and temperature dependence of the glassy dynamics. Furthermore, we determine a crossover temperature using independent dynamical and thermodynamic criteria and compare with the theoretically predicted crossover temperature $T_a$. Relaxation dynamics is shown to obey density-temperature scaling, similar to thermodynamic properties, in terms of a variable $Γ$ formed by an appropriate combination of density and temperature, characteristic of IPL interactions. Finally, we show that, when the excess thermodynamic and dynamic quantities obtained at different densities are plotted as functions of the reduced temperature $T/T_a$ (or $T_a/T$), the data collapse onto master curves with excellent agreement between theory and simulation. These scaling relations provide a unified description of the thermodynamics and dynamics in IPL systems, enabling the prediction of relaxation behavior over a wide range of densities from data at a single state point.

cond-mat.soft↗

arXiv2026

Enabling Cosmic Web Analysis at Gigaparsec Scales: A Multi Block Approach for DisPerSE

Cosmic filaments are the longest structures in the Universe and the dominant element of the cosmic web, channelling matter onto clusters and shaping the environments in which galaxies form and evolve. Accurate reconstructions of this network across gigaparsec volumes are increasingly important for cosmology and galaxy evolution. However, the most commonly used topological filament finder, DisPerSE (Discrete Persistent Structures Extractor), faces a memory bottleneck: it requires a Delaunay tessellation of the full input point set, preventing application to large simulations. Naively splitting the volume fails, as different sub-volumes yield inconsistent tessellations and filament networks. We present a frozen-core method that overcomes this bottleneck while preserving the global topology. The volume is decomposed into overlapping blocks whose tessellations are filtered by a circumsphere criterion retaining only globally valid tetrahedra; a post-processing pipeline merges the tiled outputs through core filtering, deduplication, and boundary stitching. Validation against a monolithic reference on a $300\,h^{-1}\,\mathrm{Mpc}$ MDPL2 subvolume shows 99.6% total length recovery, 100% recovery of density maxima and minima, and 94.7% individual filament matching (the ${\sim}$5% of unmatched filaments are predominantly short, low-significance structures). We apply the method to the full $(1\,h^{-1}\,\mathrm{Gpc})^3$ MDPL2 box (92 million haloes), producing a gigaparsec-scale filament catalogue. As a first application, we measure the connectivity ($κ$) for 22,900 haloes spanning $M_{200\mathrm{c}} = 10^{12}$-$10^{15.5}\,h^{-1}\,\mathrm{M}_\odot$, finding a power-law mass-connectivity relation that extends from group to cluster scales, providing the first confirmation in an $N$-body halo catalogue that the theoretically predicted scaling holds across three decades in halo mass.

astro-ph.GA↗

arXiv2026

Microscopic constitutive theory of stress overshoot, yielding, and strain hardening in amorphous materials

We develop a microscopic constitutive theory for the nonlinear deformation of metallic and polymer glasses based on nonaffine elasticity coupled to irreversible many-body relaxation. The theory predicts the full stress--strain response, from linear elasticity through stress overshoot and yielding to steady plastic flow. We show that stress overshoot originates from the competition between a nonaffine elastic instability induced by strain-driven loss of mechanical connectivity at the atomic/molecular level, and viscous dissipation associated with structural relaxation. For polymer glasses, finite chain extensibility naturally accounts for strain hardening at large deformation. The stretched-exponential relaxation exponent is obtained independently from stress or modulus relaxation measurements and provides the primary dynamical input to the theory. Using a small set of physically meaningful parameters, the model quantitatively reproduces experimental stress--strain curves for metallic glasses, polycarbonate, PMMA, and epoxy resins over a broad range of strain rates. These results establish a unified microscopic framework linking relaxation dynamics, yielding, plastic flow, and strain hardening in amorphous solids.

cond-mat.soft↗

arXiv2026

Sharp Spectral Bounds for Symmetric Positive Definite Tensors via Multiple Algebraic Invariants

We extend the trace--determinant framework of Nayak, Sharma, and Mishra~\cite{nayak2026} for bounding the H-eigenvalues of symmetric positive definite tensors. First, we replace the Arithmetic--Geometric Mean (AM--GM) relaxation underlying previous bounds by the exact solution of the associated constrained optimization problem, yielding sharp upper and lower bounds that are attained on the admissible spectral variety. Second, we incorporate higher-order power sums as additional spectral invariants and prove a structural theorem showing that any extremizer over a $K$-invariant feasibility region has at most $K$ distinct spectral values. This reduces the problem to a finite collection of low-dimensional polynomial systems and yields a hierarchy of increasingly tight bounds. For the four-invariant case $(T,S,p_3,D)$, we develop a complete theory including solution-count estimates, a multistart Newton algorithm, and sharpness conditions. We also derive closed-form bounds in small dimensions, establish perturbation estimates, and obtain refined Lyapunov region-of-attraction bounds. Numerical experiments for dimensions up to $d=100$ show that the sharp three-invariant bound reduces the median relative overestimation gap from $53\%$ to $6\%$ while maintaining low computational cost. The framework is validated on tensors with real H-spectrum.

math.OC↗

arXiv2026

SigLino: Efficient Multi-Teacher Distillation for Agglomerative Vision Foundation Models

Vision foundation models trained via multi-teacher distillation offer a promising path toward unified visual representations, yet the learning dynamics and data efficiency of such approaches remain underexplored. In this paper, we systematically study multi-teacher distillation for vision foundation models and identify key factors that enable training at lower computational cost. We introduce SigLino, an efficient family of agglomerative vision foundation models that distill knowledge from SigLIP2 and DINOv3 simultaneously into Dense and Mixture-of-Experts students. We show that (1) our Asymmetric Relation-Knowledge Distillation loss preserves the geometric properties of each teacher while enabling effective knowledge transfer, (2) token-balanced batching that packs varying-resolution images into sequences with uniform token budgets stabilizes representation learning across resolutions without sacrificing performance, (3) hierarchical clustering and sampling of training data, typically reserved for self-supervised learning, substantially improves sample efficiency over random sampling for multi-teacher distillation, and (4) the resulting representations transfer effectively to early-fusion Grounding-VLMs, outperforming models trained from scratch. By combining these findings, we curate OpenLVD200M, a 200M-image corpus that demonstrates superior efficiency for multi-teacher distillation. Instantiated in a Mixture-of-Experts, our SigLino-MoE initializes an early-fusion Grounding-VLM that replaces the conventional ViT->LLM stack, demonstrating improved performance compared to a model trained from scratch. We release OpenLVD200M and five distilled checkpoints comprising MoE and dense variants.

cs.CV↗

arXiv2026

Falcon Perception

Perception-centric systems are typically implemented with a modular encoder-decoder pipeline: a vision backbone for feature extraction and a separate decoder (or late-fusion module) for task prediction. This raises a central question: is this architectural separation essential or can a single early-fusion stack do both perception and task modeling at scale? We introduce Falcon Perception, a unified dense Transformer that processes image patches and text tokens in a shared parameter space from the first layer, using a hybrid attention pattern (bidirectional among image tokens, causal for prediction tokens) to combine global visual context with autoregressive, variable-length instance generation. To keep dense outputs practical, Falcon Perception retains a lightweight token interface and decodes continuous spatial outputs with specialized heads, enabling parallel high-resolution mask prediction. Our design promotes simplicity: we keep a single scalable backbone and shift complexity toward data and training signals, adding only small heads where outputs are continuous and dense. On SA-Co, Falcon Perception improves mask quality to 68.0 Macro-F$_1$ compared to 62.3 of SAM3. We also introduce PBench, a benchmark targeting compositional prompts (OCR, spatial constraints, relations) and dense long-context regimes, where the model shows better gains. Finally, we extend the same early-fusion recipe to Falcon OCR: a compact 300M-parameter model which attains 80.3% on olmOCR and 88.64 on OmniDocBench.

cs.CV↗

arXiv2026

Atomistic Framework for Glassy Polymer Viscoelasticity Across Twenty Frequency Decades

Glassy polymers are central to engineering applications, yet their viscoelastic response over broad frequency and temperature ranges remains difficult to characterize. We extend non-affine deformation theory by incorporating a time-dependent memory kernel within the Generalized Langevin Equation for atomistic non-affine motions, yielding frequency-dependent mechanical response. Applied to poly(methyl methacrylate) (PMMA), the method captures the shear modulus and relaxation spectrum across more than twenty decades in frequency, from hundreds of terahertz to the millihertz regime, thus bridging polymer mechanics from ordinary to extreme scales. Our predictions show quantitative consistency with independent estimates from oscillatory-shear molecular dynamics, Brillouin scattering, ultrasonic spectroscopy, Split-Hopkinson testing, and dynamic mechanical analysis (DMA), demonstrating a unified theoretical-computational route for multiscale characterization of polymer glasses.

cond-mat.soft↗

arXiv2026

Unified Learning-to-Rank for Multi-Channel Retrieval in Large-Scale E-Commerce Search

Large-scale e-commerce search must surface a broad set of items from a vast catalog, ranging from bestselling products to new, trending, or seasonal items. Modern systems therefore rely on multiple specialized retrieval channels to surface products, each designed to satisfy a specific objective. A key challenge is how to effectively merge documents from these heterogeneous channels into a single ranked list under strict latency constraints while optimizing for business KPIs such as user conversion. Rank-based fusion methods such as Reciprocal Rank Fusion (RRF) and Weighted Interleaving rely on fixed global channel weights and treat channels independently, failing to account for query-specific channel utility and cross-channel interactions. We observe that multi-channel fusion can be reformulated as a query-dependent learning-to-rank problem over heterogeneous candidate sources. In this paper, we propose a unified ranking model that learns to merge and rank documents from multiple retrieval channels. We formulate the problem as a channel-aware learning-to-rank task that jointly optimizes clicks, add-to-carts, and purchases while incorporating channel-specific objectives. We further incorporate recent user behavioral signals to capture short-term intent shifts that are critical for improving conversion in multi-channel ranking. Our online A/B experiments show that the proposed approach outperforms rank-based fusion methods, leading to a +2.85\% improvement in user conversion. The model satisfies production latency requirements, achieving a p95 latency of under 50\,ms, and is deployed on Target.com.

cs.IR↗

arXiv2026

Atrial Fibrillation Detection Using Machine Learning

Atrial fibrillation (AF) is a common cardiac arrhythmia and a major risk factor for ischemic stroke. Early detection of AF using non-invasive signals can enable timely intervention. In this work, we present a comprehensive machine learning framework for AF detection from simultaneous photoplethysmogram (PPG) and electrocardiogram (ECG) signals. We partitioned continuous recordings from 35 subjects into 525 segments (15 segments of 10,000 samples each at 125Hz per subject). After data cleaning to remove segments with missing samples, 481 segments remained (263 AF, 218 normal). We extracted 22 features per segment, including time-domain statistics (mean, standard deviation, skewness, etc.), bandpower, and heart-rate variability metrics from both PPG and ECG signals. Three classifiers -- ensemble of bagged decision trees, cubic-kernel support vector machine (SVM), and subspace k-nearest neighbors (KNN) -- were trained and evaluated using 10-fold cross-validation and hold-out testing. The subspace KNN achieved the highest test accuracy (98.7\%), slightly outperforming bagged trees (97.9\%) and cubic SVM (97.1\%). Sensitivity (AF detection) and specificity (normal rhythm detection) were all above 95\% for the top-performing models. The results indicate that ensemble-based machine learning models using combined PPG and ECG features can effectively detect atrial fibrillation. A comparative analysis of model performance along with strengths and limitations of the proposed framework is presented.

cs.CY↗

arXiv2026

The Role of Large-Scale Environment in Shaping the Stellar Mass-Gas Metallicity Relation Across Time

We study the stellar mass-gas metallicity relation (MZR) which shows a significant scatter for a fixed stellar mass. By defining global environments, nodes, filaments, and voids within the Horizon Run 5 cosmological hydrodynamical simulation, we explore when and where the enrichment of galaxies occurs, analysing key evolution parameters such as star-formation rate and changes in gas-fraction and gas-metallicity per unit time. At high redshift ($z>4.5$), there are minimal deviations from the MZR due to environment, however, larger deviations emerge as redshift decreases. Low stellar mass galaxies in nodes, $M_{\star} < 10^{9.8}\,\text{M}_{\odot}$, start showing deviations at $z = 3.5$, whilst other environments do not. For, $z < 2$, filaments and voids begin to show deviations above and below the MZR, respectively. By $z = 0.625$, the last epoch of HR5, deviations exist for all stellar masses and environments, with a maximum value of 0.13 dex at $M_{\star} \approx 10^{9.35}\,\text{M}_{\odot}$, between the median gas metallicities of node and void galaxies. To explain this environmental variance we discuss gas accretion, AGN, ram-pressure-stripping and strangulation as regulators of $Z_{g}$. Concurrently, at high metallicities, for $z < 2$, while massive galaxies in nodes show increasing $Z_{g}$ and decreasing [O/Fe], void galaxies show a turnover where $Z_{g}$ falls with decreasing [O/Fe]. This directly points to the importance of cold-gas accretion in retaining lower $Z_{g}$ in massive void galaxies for $z < 2$, whilst its absence in nodes allowed $Z_{g}$ to access higher values.

astro-ph.GA↗

arXiv2025

VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs

Vision-Language Models (VLMs) have achieved remarkable progress across tasks such as visual question answering and image captioning. Yet, the extent to which these models perform visual reasoning as opposed to relying on linguistic priors remains unclear. To address this, we introduce VisRes Bench, a benchmark designed to study visual reasoning in naturalistic settings without contextual language supervision. Analyzing model behavior across three levels of complexity, we uncover clear limitations in perceptual and relational visual reasoning capacities. VisRes isolates distinct reasoning abilities across its levels. Level 1 probes perceptual completion and global image matching under perturbations such as blur, texture changes, occlusion, and rotation; Level 2 tests rule-based inference over a single attribute (e.g., color, count, orientation); and Level 3 targets compositional reasoning that requires integrating multiple visual attributes. Across more than 19,000 controlled task images, we find that state-of-the-art VLMs perform near random under subtle perceptual perturbations, revealing limited abstraction beyond pattern recognition. We conclude by discussing how VisRes provides a unified framework for advancing abstract visual reasoning in multimodal research.

cs.CV↗

arXiv2025

Dual Guidance Semi-Supervised Action Detection

Semi-Supervised Learning (SSL) has shown tremendous potential to improve the predictive performance of deep learning models when annotations are hard to obtain. However, the application of SSL has so far been mainly studied in the context of image classification. In this work, we present a semi-supervised approach for spatial-temporal action localization. We introduce a dual guidance network to select better pseudo-bounding boxes. It combines a frame-level classification with a bounding-box prediction to enforce action class consistency across frames and boxes. Our evaluation across well-known spatial-temporal action localization datasets, namely UCF101-24 , J-HMDB-21 and AVA shows that the proposed module considerably enhances the model's performance in limited labeled data settings. Our framework achieves superior results compared to extended image-based semi-supervised baselines.

cs.CV↗