#01Sep 8, 2026
cs.CV
Canonical Color as a Lens into Concept Decodability in Vision Encoders and VLMs
Xiaofu Chen, Stella Frank, Yova Kementchedjhieva
Visual encoders construct a representation of the image input for Vision-Language models. How much conceptual, as opposed to immediately visible, information does this representation contain? We use canonical color as a controlled test case to ask whether vision encoders make canonical-color information linearly accessible, even when color is removed from the input image. We construct a dataset of objects with canonical colors, and probe vision encoders for both color and object identity using color and grayscale images. We find that canonical color remains decodable from grayscale images, and is tied to predicted object identity, indicating a conceptual link. Extending this analysis to full VLMs, we find that VLM post-training can have a surprisingly large effect on color decodability in the vision encoder. Overall, canonical color provides a usefully controllable lens for tracing object-level conceptual semantic information in vision encoders and VLMs.
#02Sep 8, 2026
cs.CV
Beyond Gait: Person Identification from Millimeter-Wave Point Clouds Across Activities of Daily Living
Xilai Wang, Zixiong Han, Saad Rhanmouni and 3 more
Person identification from millimeter-wave (mmWave) point clouds has mainly relied on gait. Indoor walking, however, is often brief and interrupted, while other activities of daily living (ADLs) may provide complementary identity information. We investigate identification across seven ADLs using mm-ADL, a new point-cloud dataset collected from 11 subjects under a controlled protocol. This extension introduces heterogeneous states and transitions whose spatial and temporal characteristics vary with activity. We therefore study whether activity can provide useful context for learning identity representations. We propose an activity-conditioned framework in which a human activity recognition router dispatches each clip to an activity-specific identity expert. The framework is implemented as a supervised mixture of experts, using a dual-stream static-dynamic PointNet (DS-SDPNet) to combine time-aggregated spatial structure with frame-to-frame information. We evaluate closed-set identification (ID) and subject-disjoint re-identification (ReID). With learned hard routing, ID accuracy increases from 62.1% to 68.0%. In a two-occupant ReID setting, hard routing increases mAP from 57.2% to 75.4% and Rank-1 accuracy from 59.1% to 82.1%. Under a matched gallery partition, activity-specific experts also outperform a shared embedding, showing that the gain extends beyond restricting the gallery. These results support the feasibility of using ADLs beyond gait for identification and the value of activity conditioning under controlled indoor conditions.
#03Sep 8, 2026
cs.CV
Studying Image Tokenizers as Visual Languages in Unified Multimodal Models
Siting Li, Zhengyang Wang, Simon Shaolei Du and 2 more
Image tokenizers define the ``visual language'' of unified multimodal models, yet are commonly studied through isolated metrics or generation-/understanding-only evaluations. These evaluations do not fully capture how visual tokens behave when modeled jointly with text. We build a controlled pure-autoregressive testbed and track task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. We examine how these losses scale and relate to downstream performance, then use them to study multimodal learnability---how well image and text tokens are jointly modeled---and tokenizer design. We find that (1) losses should be analyzed by task, since they exhibit distinct scaling behavior and rank tokenizers differently. (2) The loss--performance relationship depends on the predicted token space: for a fixed tokenizer, T2I and I2T losses correlate with generation quality, but across tokenizers, the T2I loss--performance relationship shifts with the image-token space, whereas I2T loss, computed over a shared text vocabulary, provides a more consistent signal. I2T loss also correlates with both generation and visual understanding performance after supervised finetuning. Using losses as a lens, we show that (3) better reconstruction does not necessarily yield lower task-specific losses or stronger downstream performance, and that (4) image tokenizer choice can affect text modeling under joint optimization. As case studies, we revisit three tokenizer design axes---the discriminator, semantic supervision, and vocabulary size---to examine their effects on joint modeling and downstream performance. Together, our testbed offers a complementary perspective on image tokenizers as visual languages, highlighting their interplay with text in joint multimodal training.
#04Sep 8, 2026
cs.CV
ArmPoser: Real-Time, Calibration-Free Arm Pose Estimation from Smartwatch IMU
Bishnu Dev, Vasco Xu, Xi-Aan Loh and 3 more
Arm pose estimation enables applications in fitness, extended reality input, rehabilitation, and life logging. Prior smartwatch-based approaches rely on calibration poses and preprocessing pipelines that transform raw IMU measurements into standardized training formats. These steps hinder deployment in everyday settings and introduce errors due to imperfect calibration and sensor drift. We present ArmPoser, a calibration-free arm pose estimation system using a single smartwatch IMU. Our central contribution is training models directly in the reference frame native to consumer smartwatches, aligning learning with how IMU data is produced by deployed devices. By operating on device-native axes, ArmPoser removes the need for coordinate transformations, explicit alignment, and bone-offset calibration used in prior work. We further augment training with physically grounded variations in watch placement and arm morphology to account for user-specific variability. ArmPoser also includes a wear-configuration module that infers anterior or posterior forearm placement and crown orientation. We evaluate pose estimation on public benchmarks and on a 10-participant, 30-activity study using watchOS and Android smartwatches, where ArmPoser matches or exceeds calibrated baselines without any user calibration.
#05Sep 8, 2026
cs.CV
AXS-Net: Interpretable Deep Unfolding for Hyperspectral Image Denoising via Spectral Basis Unmixing and Structured Noise Refinement
Ziyi Guan, Jianping Zhang, Zheng Yang
Hyperspectral images (HSIs) are often degraded by mixed noise, including band-dependent Gaussian perturbations and structured artifacts such as stripes, dead-lines, and impulse noise. Most deep denoisers regress the clean image directly, entangling signal and structured noise. We instead model HSI denoising as $\Y=\A\X+\Snoise+\Nnoise$, where $\A\X$ is a low-rank spectral-subspace (unmixing) reconstruction, $\Snoise$ is structured sparse noise and $\Nnoise$ is residual Gaussian noise. The resulting regularized optimization problem is unrolled into AXS-Net, a $K$-stage alternating proximal-point framework. Each stage combines an analytic spectral-basis gradient step, an SSX-Block proximal operator for abundance coefficients, and an SBlock proximal operator for the structured residual with column-consistent and sparse priors. This optimization correspondence exposes interpretable endmembers, abundance maps, and structured-noise estimates. Across ICVL, CAVE, and Harvard datasets and five noise configurations, the proposed AXS-Net achieves strong in-domain accuracy and competitive zero-shot transfer, with consistent gains across all five noise regimes on ICVL and Harvard. The recovered structured-noise closely follows the synthetic reference, and the recovered spectral basis is smooth and band-ordered rather than an arbitrary set of latent channels.