Search

Recent work showed that large diffusion models can be reused as highly precise monocular depth estimators by casting depth estimation as an image-conditional image generation task. While the proposed model achieved state-of-the-art results, high computational demands due to multi-step inference limited its use in many scenarios. In this paper, we show that the perceived inefficiency was caused by a flaw in the inference pipeline that has so far gone unnoticed. The fixed model performs comparably to the best previously reported configuration while being more than 200$\times$ faster. To optimize for downstream task performance, we perform end-to-end fine-tuning on top of the single-step model with task-specific losses and get a deterministic model that outperforms all other diffusion-based depth and normal estimation models on common zero-shot benchmarks. We surprisingly find that this fine-tuning protocol also works directly on Stable Diffusion and achieves comparable performance to current state-of-the-art diffusion-based depth and normal estimation models, calling into question some of the conclusions drawn from prior works., Comment: Project page: https://vision.rwth-aachen.de/diffusion-e2e-ft

Foundational Large Language Models (LLMs) such as GPT-3.5-turbo allow users to refine the model based on newer information, known as ``fine-tuning''. This paper leverages this ability to analyze AC-DC converter behaviors, focusing on the ripple current in DC-link capacitors. Capacitors degrade faster under high ripple currents, complicating life monitoring and necessitating preemptive replacements. Using minimal invasive noisy hardware measurements from a full bridge rectifier and 90W Power Factor Correction (PFC) boost converter, an LLM-based models to predict ripple content in DC-link currents was developed which demonstrated the LLMs' ability for near-accurate predictions. This study also highlights data requirements for precise nonlinear power electronic circuit parameter predictions to predict component degradation without any additional sensors. Furthermore, the proposed framework could be extended to any non-linear function mapping problem as well as estimating the capacitor Equivalent Series Resistance (ESR)., Comment: 6 pages, 12 figures, conference

The fibula-free flap (FFF) is a valuable reconstructive technique in maxillofacial surgery; however, the assessment of osteotomy accuracy remains challenging. We devised two novel methodologies to compare planned and postoperative osteotomies in FFF reconstructions that minimized user input but would still generalize to other operations involving the analysis of osteotomies. Our approaches leverage basic mathematics to derive both quantitative and qualitative insights about the relationship of the postoperative osteotomy to the planned model. We have coined our methods 'analysis by a shared reference angle' and 'Euler angle analysis'. In addition to describing our algorithm and the clinical utility, we present a thorough validation of both methods. Code is available at https://github.com/monkeygobah/osteoplane.

We interact with the world with our hands and see it through our own (egocentric) perspective. A holistic 3Dunderstanding of such interactions from egocentric views is important for tasks in robotics, AR/VR, action recognition and motion generation. Accurately reconstructing such interactions in 3D is challenging due to heavy occlusion, viewpoint bias, camera distortion, and motion blur from the head movement. To this end, we designed the HANDS23 challenge based on the AssemblyHands and ARCTIC datasets with carefully designed training and testing splits. Based on the results of the top submitted methods and more recent baselines on the leaderboards, we perform a thorough analysis on 3D hand(-object) reconstruction tasks. Our analysis demonstrates the effectiveness of addressing distortion specific to egocentric cameras, adopting high-capacity transformers to learn complex hand-object interactions, and fusing predictions from different views. Our study further reveals challenging scenarios intractable with state-of-the-art methods, such as fast hand motion, object reconstruction from narrow egocentric views, and close contact between two hands and objects. Our efforts will enrich the community's knowledge foundation and facilitate future hand studies on egocentric hand-object interactions., Comment: Accepted to ECCV 2024

The industrial Internet of Things (IIoT) under Industry 4.0 heralds an era of interconnected smart devices where data-driven insights and machine learning (ML) fuse to revolutionize manufacturing. A noteworthy development in IIoT is the integration of federated learning (FL), which addresses data privacy and security among devices. FL enables edge sensors, also known as peripheral intelligence units (PIUs) to learn and adapt using their data locally, without explicit sharing of confidential data, to facilitate a collaborative yet confidential learning process. However, the lower memory footprint and computational power of PIUs inherently require deep neural network (DNN) models that have a very compact size. Model compression techniques such as pruning can be used to reduce the size of DNN models by removing unnecessary connections that have little impact on the model's performance, thus making the models more suitable for the limited resources of PIUs. Targeting the notion of compact yet robust DNN models, we propose the integration of iterative magnitude pruning (IMP) of the DNN model being trained in an over-the-air FL (OTA-FL) environment for IIoT. We provide a tutorial overview and also present a case study of the effectiveness of IMP in OTA-FL for an IIoT environment. Finally, we present future directions for enhancing and optimizing these deep compression techniques further, aiming to push the boundaries of IIoT capabilities in acquiring compact yet robust and high-performing DNN models., Comment: 6 pages, 6 figures

Conventional and recently developed approaches for estimating turbulent scalar fluxes under stable conditions are evaluated. The focus is on methods that do not require fast scalar sensors such as the relaxed eddy accumulation (REA) approach, the disjunct eddy-covariance (DEC) approach, and a novel mixing length parametrization labelled as A22. Using high-frequency measurements collected from two contrasting sites (Utqiagvik, Alaska and Wendell, Idaho "during winter"), it is shown that the REA and A22 models outperform the conventional Monin-Obukhov Similarity Theory (MOST) utilized in Earth System Models. With slow trace gas sensors used in disjunct eddy-covariance (DEC) approaches and the more complex signal filtering associated with REA devices (here simulated using filtered signals from fast-response sensors), A22 outperforms REA and DEC in predicting the observed unfiltered (total) eddy-covariance (EC) fluxes. However, REA and DEC can still capture the observed filtered EC fluxes computed with the filtered scalar signal. This finding motivates the development of a correction, blending the REA and DEC methods, for the underestimated net averaged fluxes to incorporate the effect of sensor filtering. The only needed parameter for this correction is the mean velocity at the instrument height, a surrogate of the advective timescale., Comment: 14 pages, 8 figures, 2 tables

Unsteady land-sea breezes (LSBs) resulting from time-varying surface thermal contrasts are explored in the presence of a constant synoptic pressure forcing, Mg, when the latter is oriented from sea to land versus land to sea. Large eddy simulations reveal the development of four distinctive regimes depending on the joint interaction between (Mg, orientation) and thermal contrasts in modulating the fine-scale dynamics. Time lags, computed as the shifts that maximize correlation coefficients of the dynamics between transient and the corresponding steady state scenarios at maximum thermal contrast, are found to be significant and to extend 2 hours longer for sea to land compared to land to sea. These diurnal dynamics result in non-equilibrium flows that behave differently over the two patches for both orientations. Turbulence is found to be out of equilibrium with the mean flow, and the mean itself is found to be out of equilibrium with the thermal forcing. The sea surface heat flux is consistently more sensitive than its land counterpart to the time-varying external forcing, and more so for synoptic forcing from land-to-sea. Hence, although the land reaches equilibrium faster, the sea patch is found to exert a stronger control on the final turbulence-mean flow equilibrium response. Finally, vertical velocity profile at the shore and shore-normal velocity transects at the first grid level are shown to encode the multiscale regimes of the LSBs evolution, and can thus be used to identify these regimes using k-means clustering., Comment: 41 pages, 10 figures

Ensuring adequate ventilation of exterior and interior urban spaces is essential for the safety and comfort of inhabitants. Here, we examine how angled features can steer wind into areas with stagnant air, promoting natural ventilation. Using Large Eddy Simulations (LES) and wind tunnel experiments with particle image velocimetry (PIV) measurements, we first examine how louvers, located at the top of a box enclosed on four sides, can improve ventilation in the presence of incoming wind. By varying louver scale, geometry, and angle, we identify a geometric regime wherein louvers capture free-stream air to create sweeping interior flow structures, increasing the Air Exchange Rate (ACH) significantly above that for an equivalent box with an open top. We then show that non-homogeneous louver orientations enhance ventilation, accommodating winds from opposing directions, and address the generalization to taller structures. Finally, we demonstrate the feasibility of replacing louvers with lattice-cut kirigami ("cut paper"), which forms angled chutes when stretched in one direction, and could provide a mechanically preferable solution for adaptive ventilation. Our findings for this idealized system may inform the design of retrofits for urban structures -- e.g. canopies above street canyons, and "streeteries" or parklets -- capable of promoting ventilation, while simultaneously providing shade., Comment: 16 pages, 11 figures. Appendix: 2 pages, 1 figure