This year, multimodal foundation models were the star attraction. These are large, pre-trained networks that bridge computer vision and natural language processing using text prompts to query and interface with visual information. Top technology companies including Waymo, Netflix, and Meta showcased impressive advances for applications like autonomous driving, screen reconstruction, autonomous robotics, synthetic data generation, and even data curation, underscoring how mainstream the technology has become.

Just as striking was the push to shrink these behemoths for constrained hardware. Efficient large vision models (eLVMs) and FastVLM cut latency enough to run on a modest graphics processing unit (GPU) tucked inside a vehicle or tactical radio. Meanwhile, researchers showcased vision-language-action (VLA) models that skip the “text-answer” step and go straight to a robot command—think a warehouse bot that hears “Fetch the damaged pallet” and navigates there autonomously.

Overall, this field is moving beyond tasks like detection and segmentation toward more complex tasks like scene understanding and synthesis, with foundation models playing a central role. This was evident in competition challenges like those presented by Waymo/Argo, which are centered on scene understanding with tasks such as end-to-end driving and scene generation, underscoring the shift away from purely perception-based challenges.

Advances in 3D reconstruction, which is the use of two-dimensional imagery, videos, and sensor data to create 3D models, were also noteworthy, including:

• Visual Geometry Grounded Transformer (VGGT): Recognized with the conference’s best paper award, VGGT is a large neural network that generates comprehensive 3D scene information, including full camera poses, depth maps, and point clouds, in under a second using as few as one frame. Instead of relying heavily on post-processing, VGGT directly infers 3D structure using a novel transformer architecture to outperform state-of-the-art alternatives. The pretrained model is available under a permissive license, providing a fast, reliable foundation for modern 3D vision applications and significantly lowering the barrier to entry into 3D reconstruction.
• Multi-View Network for Stereo 3D Reconstruction (MUSt3R): MUSt3R is a high-performance, general-purpose method for scaling multi-view stereo correspondence to thousands of frames without external structure-from-motion preprocessing. This works to ensure that reconstructed scenes align across multiple viewpoints.
• Real-Time Dense SLAM with 3D Reconstruction Priors (MASt3R-SLAM): Unlike traditional monocular simultaneous localization and mapping (SLAM) systems that struggle with drift and scale ambiguity, MASt3R-SLAM employs learned 3D priors and differentiable rendering to refine its map and pose estimates, enabling loop detection, loop closure, and accurate dense reconstructions. This method stands out for its robustness in unconstrained settings and its ability to generate high-fidelity geometry while maintaining real-time performance.

Another key focus was recent advances in Gaussian splatting, a rendering trick that can turn casual phone footage into navigable 3D scenes in minutes. Two CVPR papers stood out in terms of this trend:

• MegaSAM solved the nagging calibration step: upload video and get a textured model, with no camera intrinsics required.
• Student Splatting and Scooping (SSS) swapped each Gaussian blob for a Student’s-t distribution, a more flexible basis function that allows users to squeeze the same realism from a fraction of the compute. The authors open-sourced the code under a permissive license.

These approaches are important, as digital twins once demanded expensive LiDAR scanners and long post-processing queues. For agencies building digital twins of bases, disaster sites, or archaeological digs, the barrier to entry may drop to “own a smartphone,” as a field team will soon be able to capture, upload, and brief leadership before wheels-up.

Conventional frame-based cameras struggle with scenes that are both bright and fast—rocket launches, muzzle flashes, hypersonic test flights. Event-based cameras sidestep the problem by recording only changes in intensity at micro-second cadence. The catch has been data formats incompatible with mainstream vision tooling.

Addressing this issue, ETH Zurich’s Davide Scaramuzza and collaborators showcased field-programmable gate array (FPGA) accelerators and “event-to-frame” accumulation tricks that slot neatly into existing deep-learning pipelines while preserving sub-5-millisecond latency. Demos tracked micro-drones in cluttered rooms and guided quadcopters through fireworks without bloom or blur.

Qualcomm’s Low-Power Computer Vision Challenge drew more than 100 teams racing to compress deep nets for micro-joule budgets. The winners achieved state-of-the-art image classification in 1 millisecond on phone-class silicon, and open-vocabulary segmentation in under 20 milliseconds. That same spirit drove a stream of talks on sparsity, quantization, and FPGA pipelines for VLMs.

For federal missions, running advanced perception models in a disconnected or contested environment is no longer optional. Booz Allen’s Vision AI stack was built for exactly that challenge, providing a secure, modular path to deploy and monitor computer-vision workloads from cloud to tactical edge.

Some of the most talked-about releases—Molmo and PixMo—shipped all code, weights, and training data under permissive licenses. This counters the “black-box foundation model” trend and aligns with federal mandates for transparency and reproducibility. Meanwhile, CVPR panels on “Vision for Safety and Security” echoed our emphasis on responsible, secure AI adoption.