Research RadarFace DetectionarXivOctober 2026

Monthly arXiv Radar

October 2026 Face Detection Papers: Lightweight Fusion, Online Speakers, and Real CCTV

October produced one direct detector paper and two useful system-level studies around localization. WeaveFace focuses on small and occluded faces under tight compute budgets. CAST-ASD shows how tracked faces can be compared online with no future frames, while CVBL-CCTV-Face tests what happens after detection when real surveillance degradation reaches enhancement and matching.

What This Month Signals

WeaveFace interleaves adjacent feature scales before a Mamba scan and reaches 91.41% mean AP on WIDER FACE with 0.34 million parameters. CAST-ASD takes detected face tracks as input and uses zero-lookahead attention across current speakers, reducing competing-speaker false alarms by 32% on MSDWild. CVBL-CCTV-Face contributes 223 identities captured under authentic surveillance degradation and finds that synthetic-degradation fine-tuning helps recognition more than direct low-quality fine-tuning. The combined lesson is that detector quality, online scene reasoning, and downstream recognition utility must be tested as one pipeline.

Paper 012026-10-05cs.CV

Weave Mamba Fusion: Global Cross-Scale Interaction for Lightweight Face Detection

arXiv PDF

Authors & Institutions

Dohun Kim

School of Data Science, The University of Suwon

Jinmyung Jung

School of Data Science, The University of Suwon

What Problem It Solves

WeaveFace aims to add global cross-scale interaction to a sub-million-parameter face detector without paying the memory and compute cost of concatenating complete feature maps or applying quadratic self-attention.

Key Result

On WIDER FACE, WeaveFace reports 91.41% mean AP with 0.34 million parameters and 1.16 GFLOPs. Its Hard-subset AP is 87.14%, the area where small and occluded faces most stress the cross-scale design. The paper reports stronger accuracy than prior detectors below 0.5 million parameters.

Abstract

Feature pyramid methods, from FPN to BiFPN, have achieved strong performance in face detection by fusing multi-scale features. However, detecting faces under unconstrained conditions, such as small scale, occlusion, and extreme pose, remains difficult, as it requires global cross-scale dependencies that local fusion cannot model. State space models such as Mamba provide global context with linear complexity by scanning features as a sequence, and therefore offer a promising direction for this problem. Nevertheless, such a scan needs the two pyramid scales combined into a single feature map, and the way they are combined determines whether cross-scale structure is preserved. Summation collapses the two scales before the scan, so the scan has no cross-scale structure to exploit, while concatenation keeps both scales but at far higher cost. To address this, we propose Weave Mamba Fusion (WMF), which interleaves two adjacent pyramid scales column by column so that each step of a horizontal bidirectional SS2D scan moves from one scale to the other. With partial-channel processing and parameter-free de-weaving, WMF enables efficient cross-scale interaction while preserving feature structure. Integrating WMF into every fusion node yields WeaveBiFPN, the neck of our WeaveFace detector. On WIDER FACE, WeaveFace achieves 91.41% mean AP with only 0.34M parameters and 1.16 GFLOPs, outperforming prior detectors under 0.5M parameters. Its largest gains are on the Hard subset, where it reaches 87.14% AP. The code is publicly available at https://github.com/dohun-mat/WeaveMambaFusion.

Research Starting Point

Small, occluded, and strongly rotated faces need fine spatial detail from high-resolution pyramid levels and semantic context from low-resolution levels. Standard FPN and BiFPN fusion is local, while full attention can be too expensive at the resolutions used by a lightweight detector.

Method

Weave Mamba Fusion interleaves columns from adjacent pyramid levels so a bidirectional horizontal SS2D scan alternates between scales at each step. Only part of the channels are processed, and a parameter-free de-weaving operation restores the original scales. Replacing each BiFPN fusion node with this operator produces WeaveBiFPN, which is used as the neck of the WeaveFace detector.

Paper Summary

This is a promising detector for edge deployments where parameter count and FLOPs matter, particularly when tiny faces dominate. Teams should still measure wall-clock latency and memory on their own accelerator because the paper's main efficiency evidence is model complexity.

Paper 022026-10-10cs.CV

CAST-ASD: Causal Attention over Speakers and Time for Online Active Speaker Detection

arXiv PDF

Authors & Institutions

Pratham Gupta

Indian Institute of Technology Kanpur

KTH Royal Institute of Technology

Haotian Qi

KTH Royal Institute of Technology

Tushar Sandhan

Indian Institute of Technology Kanpur

Gabriel Skantze

KTH Royal Institute of Technology

What Problem It Solves

CAST-ASD brings cross-face reasoning to a zero-lookahead detector whose parameter count does not grow with the number of visible faces. It addresses the online stage after face detection and tracking, where low latency and speaker attribution are both required.

Key Result

CAST-ASD reduces competing-speaker false alarms by 32% relative to its causal per-face backbone and outperforms all causal baselines on both MSDWild validation splits. One CAST block uses 1.57 million parameters and roughly one tenth of the inter-speaker parameters of a three-block fixed-roster convolution. Its causal penalty is 1.14 mAP versus 2.67 mAP for the matched LoCoNet configuration.

Abstract

We present CAST-ASD (Causal Attention over Speakers and Time), a causal joint model for active speaker detection (ASD). Recent ASD models improve performance by jointly processing all faces in a scene, but operate offline, with the full clip available before any frame is labelled. Low-latency causal detectors instead process each face independently. CAST-ASD combines joint face reasoning with fully causal inference. It applies scene-level attention across faces within the current frame and causal temporal attention over each face's history, enabling information sharing across speakers with zero lookahead. Its parameter count does not grow with the number of faces. On MSDWild, where most negative face-frames contain a visible competing speaker, CAST-ASD reduces competing-speaker false alarms by 32% relative to the causal per-face detector it extends, with larger gains as scene occupancy increases. It outperforms all causal baselines on both validation splits while using only one-tenth the inter-speaker parameters of a three-block fixed-roster convolution. Code and models are available at https://github.com/prathamg007/CAST-ASD.

Research Starting Point

Active-speaker detection must decide which visible face owns a shared audio track. Accurate joint models usually look into future frames, while causal systems often score each face independently and therefore confuse silent bystanders with the current speaker in crowded scenes.

Method

A causal audio-visual backbone encodes every face track and the shared audio stream. Each CAST block first applies scene-level self-attention across all faces in the current frame, then causal temporal attention along each face's past. The model is evaluated on MSDWild, where most negative face frames contain a competing visible speaker, and compared with per-face and fixed-roster causal baselines under matched training conditions.

Paper Summary

For live conferencing, diarization, and camera control, comparing all currently visible faces can recover much of the context lost by forbidding lookahead. The model assumes reliable face tracks, so detector and tracker failures remain part of the end-to-end risk.

Paper 032026-10-27cs.CV

CVBL-CCTV-Face Dataset for Real-World Face Enhancement and Recognition in Surveillance

arXiv PDF

Authors & Institutions

Manish Das

Computer Vision and Biometrics Lab, Indian Institute of Information Technology Allahabad

Ajay Kumar Yadav

Computer Vision and Biometrics Lab, Indian Institute of Information Technology Allahabad

Shiv Ram Dubey

Computer Vision and Biometrics Lab, Indian Institute of Information Technology Allahabad

Satish Kumar Singh

Computer Vision and Biometrics Lab, Indian Institute of Information Technology Allahabad

What Problem It Solves

The work contributes an unpaired CCTV-to-reference dataset and a full face detection, enhancement, and recognition pipeline for studying authentic compound degradation. It also compares whether pretrained enhancement, synthetic-degradation fine-tuning, or direct real-low-quality fine-tuning transfers best.

Key Result

The dataset contains 14,081 low-quality training images, 2,186 augmented high-quality references, and a test set with 1,581 low-quality and 528 high-quality images. Fine-tuning on synthetically degraded high-quality images produces the strongest recognition results, while direct fine-tuning on real low-quality crops does not improve over pretrained baselines. That negative result highlights the remaining mismatch between restoration objectives and uncontrolled surveillance artifacts.

Abstract

In real-world surveillance settings, facial captures from closed-circuit television systems suffer from degradations such as low resolution, motion blur, noise, and occlusions, which undermine dependable face recognition. In this study, we propose a framework for super-resolution and identity recognition, aimed at surveillance applications. For this purpose, we present the CVBL-CCTV-Face Dataset which is a dataset of real-world images taken in an unpaired manner where each pair consists of a low-quality (LQ) image of a face captured by a 8MP IP Dome CCTV camera together with a high-quality (HQ) reference image that is taken in controlled settings. In order to cope with the various problems posed by such datasets, we construct a modular system that involves face detection, deep learning-based super-resolution of low-quality face images, and identity recognition through L2-normalized embeddings extracted by an embedding model. Enhancement models are trained and tested according to three different setups, and the system is deployed for identity verification in a real-time unconstrained environment.

Research Starting Point

Surveillance face pipelines are often tested with studio images or synthetic downsampling, while actual CCTV crops combine motion blur, sensor noise, pose changes, occlusion, distance, and poor illumination. Enhancement models trained on tidy paired data may therefore produce attractive images without improving identity matching on real cameras.

Method

CVBL-CCTV-Face contains 223 identities captured with an 8 MP IP dome camera plus separately acquired high-quality references. Two versions model sparse and richer enrollment, and identity-disjoint splits support evaluation. The benchmark uses MTCNN detection, four transformer or diffusion enhancement models, normalized face embeddings, and L2 matching, then deploys the same modules in a live CCTV demonstration.

Paper Summary

The value is less a new production model than a reality check and test asset for surveillance pipelines. It shows that collecting real camera data is essential, but simply fine-tuning a restoration network on those crops may not improve identity decisions.