Anthony Fuller

Daniel G. Kyrollos, Anthony Fuller, Kim Greenwood et al.

Infant pose monitoring during sleep has multiple applications in both healthcare and home settings. In a healthcare setting, pose detection can be used for region of interest detection and movement detection for noncontact based monitoring systems. In a home setting, pose detection can be used to detect sleep positions which has shown to have a strong influence on multiple health factors. However, pose monitoring during sleep is challenging due to heavy occlusions from blanket coverings and low lighting. To address this, we present a novel dataset, Simultaneously-collected multimodal Mannequin Lying pose (SMaL) dataset, for under the cover infant pose estimation. We collect depth and pressure imagery of an infant mannequin in different poses under various cover conditions. We successfully infer full body pose under the cover by training state-of-art pose estimation methods and leveraging existing multimodal adult pose datasets for transfer learning. We demonstrate a hierarchical pretraining strategy for transformer-based models to significantly improve performance on our dataset. Our best performing model was able to detect joints under the cover within 25mm 86% of the time with an overall mean error of 16.9mm. Data, code and models publicly available at https://github.com/DanielKyr/SMaL

Anthony Fuller, Koreen Millard, James R. Green

Although the remote sensing (RS) community has begun to pretrain transformers (intended to be fine-tuned on RS tasks), it is unclear how these models perform under distribution shifts. Here, we pretrain a new RS transformer--called SatViT-V2--on 1.3 million satellite-derived RS images, then fine-tune it (along with five other models) to investigate how it performs on distributions not seen during training. We split an expertly labeled land cover dataset into 14 datasets based on source biome. We train each model on each biome separately and test them on all other biomes. In all, this amounts to 1638 biome transfer experiments. After fine-tuning, we find that SatViT-V2 outperforms SatViT-V1 by 3.1% on in-distribution (matching biomes) and 2.8% on out-of-distribution (mismatching biomes) data. Additionally, we find that initializing fine-tuning from the linear probed solution (i.e., leveraging LPFT [1]) improves SatViT-V2's performance by another 1.2% on in-distribution and 2.4% on out-of-distribution data. Next, we find that pretrained RS transformers are better calibrated under distribution shifts than non-pretrained models and leveraging LPFT results in further improvements in model calibration. Lastly, we find that five measures of distribution shift are moderately correlated with biome transfer performance. We share code and pretrained model weights. (https://github.com/antofuller/SatViT)

Anthony Fuller, Koreen Millard, James R. Green

A vital and rapidly growing application, remote sensing offers vast yet sparsely labeled, spatially aligned multimodal data; this makes self-supervised learning algorithms invaluable. We present CROMA: a framework that combines contrastive and reconstruction self-supervised objectives to learn rich unimodal and multimodal representations. Our method separately encodes masked-out multispectral optical and synthetic aperture radar samples -- aligned in space and time -- and performs cross-modal contrastive learning. Another encoder fuses these sensors, producing joint multimodal encodings that are used to predict the masked patches via a lightweight decoder. We show that these objectives are complementary when leveraged on spatially aligned multimodal data. We also introduce X- and 2D-ALiBi, which spatially biases our cross- and self-attention matrices. These strategies improve representations and allow our models to effectively extrapolate to images up to 17.6x larger at test-time. CROMA outperforms the current SoTA multispectral model, evaluated on: four classification benchmarks -- finetuning (avg. 1.8%), linear (avg. 2.4%) and nonlinear (avg. 1.4%) probing, kNN classification (avg. 3.5%), and K-means clustering (avg. 8.4%); and three segmentation benchmarks (avg. 6.4%). CROMA's rich, optionally multimodal representations can be widely leveraged across remote sensing applications.

Scott C. Lowe, Anthony Fuller, Sageev Oore et al.

The landscape of self-supervised learning (SSL) is currently dominated by generative approaches (e.g., MAE) that reconstruct raw low-level data, and predictive approaches (e.g., I-JEPA) that predict high-level abstract embeddings. While generative methods provide strong grounding, they are computationally inefficient for high-redundancy modalities like imagery, and their training objective does not prioritize learning high-level, conceptual features. Conversely, predictive methods often suffer from training instability due to their reliance on the non-stationary targets of final-layer self-distillation. We introduce Bootleg, a method that bridges this divide by tasking the model with predicting latent representations from multiple hidden layers of a teacher network. This hierarchical objective forces the model to capture features at varying levels of abstraction simultaneously. We demonstrate that Bootleg significantly outperforms comparable baselines (+10% over I-JEPA) on classification of ImageNet-1K and iNaturalist-21, and semantic segmentation of ADE20K and Cityscapes.

Anthony Fuller, Yousef Yassin, Daniel G. Kyrollos et al.

We introduce a simple enhancement of vision transformers (ViTs) to improve accuracy while maintaining throughput. Our approach, Jumbo, creates a wider CLS token, which is split to match the patch token width before attention, processed with self-attention, and reassembled. After attention, Jumbo applies a dedicated, wider FFN to this token. Since there is only one Jumbo token, its cost is minimal, and because we share this FFN across layers, its parameter count is controlled. Jumbo significantly improves over ViT+Registers on ImageNet-1K and ImageNet-21K. These gains are largest at small sizes / high speeds, e.g., ViT-nano+Jumbo outperforms ViT-nano+Registers by 13%. In fact, our Jumbo models are so efficient that they outperform specialized compute-efficient models while preserving the architectural advantages of plain ViTs, such as support for token dropping and other modalities. Accordingly, we demonstrate that Jumbo excels in these two settings via masked autoencoding and on a suite of time series benchmarks. Code and weights available: https://github.com/antofuller/jumbo

Isaac Corley, Nils Lehmann, Caleb Robinson et al.

Geospatial foundation models (GFMs) have been proposed as generalizable backbones for disaster response, land-cover mapping, food-security monitoring, and other high-stakes Earth-observation tasks. Yet the published work about these models does not give reviewers or users enough information to tell which model fits a given task. We argue that nobody knows what the current state of the art is in geospatial foundation models. The methods may be useful, but the GFM literature does not standardize evaluations, training and testing protocols, released weights, or pretraining controls well enough for anyone to compare or rank them. In a 152-paper audit, we find 46 cross-paper disagreements of at least 10 points for the same model, benchmark, and protocol; 94/126 papers with extractable pretraining data use a configuration no other paper uses; and 39% of GFM papers release no model weights. This lack of community standards can be solved. We propose six concrete expectations: named-license weight release, shared core evaluations, copied-versus-rerun baseline annotations, variance reporting, one shared evaluation harness, and data-vs-architecture-vs-algorithm controls. These gaps are a coordination failure, not a fault of any individual lab; the authors of this paper, like many others in the GFM community, have contributed to them. Rather than just critiquing the community, we aim to provide concrete steps toward a shared understanding of how to innovate GFMs.

Ali Salamatian, Anthony Fuller, Pritam Sarkar et al.

Transformers dominate video recognition. They split videos into tokens, and processing them has expensive superlinear computational cost. Yet videos are filled with redundancy, so we can question the need for this expense. We introduce LookWhen, a selector-extractor framework that factorizes video recognition into learning when, where, and what to compute. Our shallow selector gets a scaled-down video and quickly scores all tokens across space-time, while our deep extractor gets the top-K selected tokens to approximate full-video representations without actually processing all the tokens. A key challenge is defining effective supervision for selection and extraction. For selection pre-training, we introduce a score on representations that ranks tokens by uniqueness using a simple nearest-neighbor distance. For extraction pre-training, we distill both a video teacher and an image teacher, for which we normalize its frame-wise representations to learn what changes within videos. Through these strategies, our selector-extractor learns general and efficient representations for feature extraction or fine-tuning to a task. Through experiments on Kinetics-400, SSv2, Epic-Kitchens, Diving48, Jester, and Charades, we show that LookWhen achieves a better accuracy-computation trade-off than efficient models and upgraded baselines of similar size. LookWhen Pareto-dominates in accuracy-FLOPs on 9 of 12 cases (6 tasks x 2 settings) and roughly matches on 3. In accuracy-throughput, measuring time in practice, LookWhen is more efficient still at 6.7x faster than InternVideo2-B at equal accuracy.

Gabriel Tseng, Anthony Fuller, Marlena Reil et al.

We introduce a highly multimodal transformer to represent many remote sensing modalities - multispectral optical, synthetic aperture radar, elevation, weather, pseudo-labels, and more - across space and time. These inputs are useful for diverse remote sensing tasks, such as crop mapping and flood detection. However, learning shared representations of remote sensing data is challenging, given the diversity of relevant data modalities, and because objects of interest vary massively in scale, from small boats (1-2 pixels and fast) to glaciers (thousands of pixels and slow). We present a novel self-supervised learning algorithm that extracts multi-scale features across a flexible set of input modalities through masked modeling. Our dual global and local contrastive losses differ in their targets (deep representations vs. shallow input projections) and masking strategies (structured vs. not). Our Galileo is a single generalist model that outperforms SoTA specialist models for satellite images and pixel time series across eleven benchmarks and multiple tasks.

Anthony Fuller, Daniel G. Kyrollos, Yousef Yassin et al.

High-resolution images offer more information about scenes that can improve model accuracy. However, the dominant model architecture in computer vision, the vision transformer (ViT), cannot effectively leverage larger images without finetuning -- ViTs poorly extrapolate to more patches at test time, although transformers offer sequence length flexibility. We attribute this shortcoming to the current patch position encoding methods, which create a distribution shift when extrapolating. We propose a drop-in replacement for the position encoding of plain ViTs that restricts attention heads to fixed fields of view, pointed in different directions, using 2D attention masks. Our novel method, called LookHere, provides translation-equivariance, ensures attention head diversity, and limits the distribution shift that attention heads face when extrapolating. We demonstrate that LookHere improves performance on classification (avg. 1.6%), against adversarial attack (avg. 5.4%), and decreases calibration error (avg. 1.5%) -- on ImageNet without extrapolation. With extrapolation, LookHere outperforms the current SoTA position encoding method, 2D-RoPE, by 21.7% on ImageNet when trained at $224^2$ px and tested at $1024^2$ px. Additionally, we release a high-resolution test set to improve the evaluation of high-resolution image classifiers, called ImageNet-HR.

Youssef Megahed, Anthony Fuller, Saleh Abou-Alwan et al.

Hirschsprung's disease (HD) is a congenital birth defect diagnosed by identifying the lack of ganglion cells within the colon's muscularis propria, specifically within the myenteric plexus regions. There may be advantages for quantitative assessments of histopathology images of the colon, such as counting the ganglion and assessing their spatial distribution; however, this would be time-intensive for pathologists, costly, and subject to inter- and intra-rater variability. Previous research has demonstrated the potential for deep learning approaches to automate histopathology image analysis, including segmentation of the muscularis propria using convolutional neural networks (CNNs). Recently, Vision Transformers (ViTs) have emerged as a powerful deep learning approach due to their self-attention. This study explores the application of ViTs for muscularis propria segmentation in calretinin-stained histopathology images and compares their performance to CNNs and shallow learning methods. The ViT model achieved a DICE score of 89.9% and Plexus Inclusion Rate (PIR) of 100%, surpassing the CNN (DICE score of 89.2%; PIR of 96.0%) and k-means clustering method (DICE score of 80.7%; PIR 77.4%). Results assert that ViTs are a promising tool for advancing HD-related image analysis.

Anthony Fuller, James R. Green, Evan Shelhamer

Model soups are strange and strangely effective combinations of parameters. They take a model (the stock), fine-tune it into multiple models (the ingredients), and then mix their parameters back into one model (the soup) to improve predictions. While all known soups require supervised learning, and optimize the same loss on labeled data, our recipes for Self-\emph{Soup}ervision generalize soups to self-supervised learning (SSL). Our Self-Souping lets us flavor ingredients on new data sources, e.g. from unlabeled data from a task for transfer or from a shift for robustness. We show that Self-Souping on corrupted test data, then fine-tuning back on uncorrupted train data, boosts robustness by +3.5\% (ImageNet-C) and +7\% (LAION-C). Self-\emph{Soup}ervision also unlocks countless SSL algorithms to cook the diverse ingredients needed for more robust soups. We show for the first time that ingredients can differ in their SSL hyperparameters -- and more surprisingly, in their SSL algorithms. We cook soups of MAE, MoCoV3, and MMCR ingredients that are more accurate than any one single SSL ingredient.

Youssef Megahed, Saleh Abou-Alwan, Anthony Fuller et al.

Hirschsprung Disease is characterized by the absence of ganglion cells in the myenteric plexus. Therefore, the correct identification of ganglion cells is crucial for diagnosing Hirschsprung disease. We introduce a three-stage analysis framework that mimics the pathologist's diagnostic approach. The framework, based on a Vision Transformer model (ViT-B/16), sequentially segments the muscularis propria, segments the myenteric plexus, and detects ganglion cells within anatomically valid regions. 30 whole-slide images of colon tissue were used, each containing manual annotations of muscularis, plexus, and ganglion cells. A 5-fold cross-validation scheme was applied to each stage, along with resolution-specific tiling strategies and tailored postprocessing to ensure anatomical consistency. The proposed method achieved a Dice coefficient of 89.9% and a Plexus Inclusion Rate of 100% for muscularis segmentation. Plexus segmentation reached a recall of 94.8%, a precision of 84.2% and a Ganglia Inclusion Rate of 99.7%. For ganglion cells annotated with high certainty, the model achieved 62.1\% precision and 89.1% recall. When considering all annotated ganglion cells, regardless of certainty level, the overall precision was 67.0%. These results indicate that ViT-based models are effective at leveraging global tissue context and capturing cellular morphology at small scales, even within complex histological tissue structures. This multi-stage methodology has great potential to support digital pathology workflows by reducing inter-observer variability and assisting in the evaluation of Hirschsprung disease. The clinical impact will be evaluated in future work with larger multi-center datasets and additional expert annotations.

Anthony Fuller, Yousef Yassin, Junfeng Wen et al.

Vision transformers are ever larger, more accurate, and more expensive to compute. The expense is even more extreme at high resolution as the number of tokens grows quadratically with the image size. We turn to adaptive computation to cope with this cost by learning to predict where to compute. Our LookWhere method divides the computation between a low-resolution selector and a high-resolution extractor without ever processing the full high-resolution input. We jointly pretrain the selector and extractor without task supervision by distillation from a self-supervised teacher, in effect, learning where and what to compute simultaneously. Unlike prior token reduction methods, which pay to save by pruning already-computed tokens, and prior token selection methods, which require complex and expensive per-task optimization, LookWhere economically and accurately selects and extracts transferrable representations of images. We show that LookWhere excels at sparse recognition on high-resolution inputs (Traffic Signs), maintaining accuracy while reducing FLOPs by up to 34x and time by 6x. It also excels at standard recognition tasks that are global (ImageNet classification) or local (ADE20K segmentation), improving accuracy while reducing time by 1.36x.