Large Language Models (LLMs) are increasingly used in applications requiring long context lengths, but the key-value (KV) cache often becomes a memory bottleneck on GPUs as context grows. To address this, we propose Commutative Vector Quantization (CommVQ) to significantly reduce memory usage for long-context LLM inference. We first introduce additive quantization with a lightweight encoder and codebook to compress the KV cache, which can be decoded via simple matrix multiplication. To further reduce computational costs during decoding, we design the codebook to be commutative with Rotary Position Embedding (RoPE) and train it using an Expectation-Maximization (EM) algorithm. This enables efficient integration of decoding into the self-attention mechanism. Our approach achieves high accuracy with additive quantization and low overhead via the RoPE-commutative codebook. Experiments on long-context benchmarks and GSM8K show that our method reduces FP16 KV cache size by 87.5% with 2-bit quantization, while outperforming state-of-the-art KV cache quantization methods. Notably, it enables 1-bit KV cache quantization with minimal accuracy loss, allowing a LLaMA-3.1 8B model to run with a 128K context length on a single RTX 4090 GPU. The source code is available at: https://github.com/UMass-Embodied-AGI/CommVQ.

On-device machine learning (ML) moves computation from the cloud to personal devices, protecting user privacy and enabling intelligent user experiences. However, fitting models on devices with limited resources presents a major technical challenge: practitioners need to optimize models and balance hardware metrics such as model size, latency, and power. To help practitioners create efficient ML models, we designed and developed Talaria : a model visualization and optimization system. Talaria enables practitioners to compile models to hardware, interactively visualize model statistics, and simulate optimizations to test the impact on inference metrics. Since its internal deployment two years ago, we have evaluated Talaria using three methodologies: (1) a log analysis highlighting its growth of 800+ practitioners submitting 3,600+ models; (2) a usability survey with 26 users assessing the utility of 20 Talaria features; and (3) a qualitative interview with the 7 most active users about their experience using Talaria.

Real world images often have highly imbalanced content density. Some areas are very uniform, e.g., large patches of blue sky, while other areas are scattered with many small objects. Yet, the commonly used successive grid downsampling strategy in convolutional deep networks treats all areas equally. Hence, small objects are represented in very few spatial locations, leading to worse results in tasks such as segmentation. Intuitively, retaining more pixels representing small objects during downsampling helps to preserve important information. To achieve this, we propose AutoFocusFormer (AFF), a local-attention transformer image recognition backbone, which performs adaptive down-sampling by learning to retain the most important pixels for the task. Since adaptive downsampling generates a set of pixels irregularly distributed on the image plane, we abandon the classic grid structure. Instead, we develop a novel point-based local attention block, facilitated by a balanced clustering module and a learnable neighborhood merging module, which yields representations for our point-based versions of state-of-the-art segmentation heads. Experiments show that our AutoFocusFormer (AFF) improves significantly over baseline models of similar sizes.

As neural networks grow in size and complexity, inference speeds decline. To combat this, one of the most effective compression techniques -- channel pruning -- removes channels from weights. However, for multi-branch segments of a model, channel removal can introduce inference-time memory copies. In turn, these copies increase inference latency -- so much so that the pruned model can be slower than the unpruned model. As a workaround, pruners conventionally constrain certain channels to be pruned together. This fully eliminates memory copies but, as we show, significantly impairs accuracy. We now have a dilemma: Remove constraints but increase latency, or add constraints and impair accuracy. In response, our insight is to reorder channels at export time, (1) reducing latency by reducing memory copies and (2) improving accuracy by removing constraints. Using this insight, we design a generic algorithm UPSCALE to prune models with any pruning pattern. By removing constraints from existing pruners, we improve ImageNet accuracy for post-training pruned models by 2.1 points on average -- benefiting DenseNet (+16.9), EfficientNetV2 (+7.9), and ResNet (+6.2). Furthermore, by reordering channels, UPSCALE improves inference speeds by up to 2x over a baseline export.

Reconstructing an accurate 3D object model from a few image observations remains a challenging problem in computer vision. State-of-the-art approaches typically assume accurate camera poses as input, which could be difficult to obtain in realistic settings. In this paper, we present FvOR, a learning-based object reconstruction method that predicts accurate 3D models given a few images with noisy input poses. The core of our approach is a fast and robust multi-view reconstruction algorithm to jointly refine 3D geometry and camera pose estimation using learnable neural network modules. We provide a thorough benchmark of state-of-the-art approaches for this problem on ShapeNet. Our approach achieves best-in-class results. It is also two orders of magnitude faster than the recent optimization-based approach IDR [67].

Deep learning has made significant impacts on multiview stereo systems. State-of-the-art approaches typically involve building a cost volume, followed by multiple 3D convolution operations to recover the input image's pixel-wise depth. While such end-to-end learning of plane-sweeping stereo advances public benchmarks' accuracy, they are typically very slow to compute. We present MVS2D, a highly efficient multi-view stereo algorithm that seamlessly integrates multi-view constraints into single-view net-works via an attention mechanism. Since MVS2D only builds on 2D convolutions, it is at least$2\times faster$than all the notable counterparts. Moreover, our algorithm produces precise depth estimations and 3D reconstructions, achieving state-of-the-art results on challenging benchmarks ScanNet, SUN3D, RGBD, and the classical DTU dataset. our algorithm also outperforms all other algorithms in the setting of inexact camera poses. Our code is released at https://github.com/zhenpeiyang/MVS2D

We study the task of semantic mapping – specifically, an embodied agent (a robot or an egocentric AI assistant) is given a tour of a new environment and asked to build an allocentric top-down semantic map (‘what is where?’) from egocentric observations of an RGB-D camera with known pose (via localization sensors). Importantly, our goal is to build neural episodic memories and spatio-semantic representations of 3D spaces that enable the agent to easily learn subsequent tasks in the same space – navigating to objects seen during the tour (‘Find chair’) or answering questions about the space (‘How many chairs did you see in the house?’). Towards this goal, we present Semantic MapNet (SMNet), which consists of: (1) an Egocentric Visual Encoder that encodes each egocentric RGB-D frame, (2) a Feature Projector that projects egocentric features to appropriate locations on a floor-plan, (3) a Spatial Memory Tensor of size floor-plan length×width×feature-dims that learns to accumulate projected egocentric features, and (4) a Map Decoder that uses the memory tensor to produce semantic top-down maps. SMNet combines the strengths of (known) projective camera geometry and neural representation learning. On the task of semantic mapping in the Matterport3D dataset, SMNet significantly outperforms competitive baselines by 4.01−16.81% (absolute) on mean-IoU and 3.81−19.69% (absolute) on Boundary-F1 metrics. Moreover, we show how to use the spatio-semantic allocentric representations build by SMNet for the task of ObjectNav and Embodied Question Answering. Project page: https://vincentcartillier.github.io/smnet.html.

We develop new representations and algorithms for three-dimensional (3D) object detection and spatial layout prediction in cluttered indoor scenes. We first propose a clouds of oriented gradient (COG) descriptor that links the 2D appearance and 3D pose of object categories, and thus accurately models how perspective projection affects perceived image gradients. To better represent the 3D visual styles of large objects and provide contextual cues to improve the detection of small objects, we introduce latent support surfaces. We then propose a "Manhattan voxel" representation which better captures the 3D room layout geometry of common indoor environments. Effective classification rules are learned via a latent structured prediction framework. Contextual relationships among categories and layout are captured via a cascade of classifiers, leading to holistic scene hypotheses that exceed the state-of-the-art on the SUN RGB-D database.

We tackle the problem of automatically reconstructing a complete 3D model of a scene from a single RGB image. This challenging task requires inferring the shape of both visible and occluded surfaces. Our approach utilizes viewer-centered, multi-layer representation of scene geometry adapted from recent methods for single object shape completion. To improve the accuracy of view-centered representations for complex scenes, we introduce a novel "Epipolar Feature Transformer" that transfers convolutional network features from an input view to other virtual camera viewpoints, and thus better covers the 3D scene geometry. Unlike existing approaches that first detect and localize objects in 3D, and then infer object shape using category-specific models, our approach is fully convolutional, end-to-end differentiable, and avoids the resolution and memory limitations of voxel representations. We demonstrate the advantages of multi-layer depth representations and epipolar feature transformers on the reconstruction of a large database of indoor scenes.

Passive visual systems typically fail to recognize objects in the amodal setting where they are heavily occluded. In contrast, humans and other embodied agents have the ability to move in the environment and actively control the viewing angle to better understand object shapes and semantics. In this work, we introduce the task of Embodied Amodel Recognition (EAR): an agent is instantiated in a 3D environment close to an occluded target object, and is free to move in the environment to perform object classification, amodal object localization, and amodal object segmentation. To address this problem, we develop a new model called Embodied Mask R-CNN for agents to learn to move strategically to improve their visual recognition abilities. We conduct experiments using a simulator for indoor environments. Experimental results show that: 1) agents with embodiment (movement) achieve better visual recognition performance than passive ones and 2) in order to improve visual recognition abilities, agents can learn strategic paths that are different from shortest paths.

Convolutional neural networks process input data by sending channel-wise feature response maps to subsequent layers. While a lot of progress has been made by making networks deeper, information from each channel can only be propagated from lower levels to higher levels in a hierarchical feed-forward manner. When viewing each filter in the convolutional layer as a neuron, those neurons are not communicating explicitly within each layer in CNNs. We introduce a novel network unit called Cross-channel Communication (C3) block, a simple yet effective module to encourage the neuron communication within the same layer. The C3 block enables neurons to exchange information through a micro neural network, which consists of a feature encoder, a message communicator, and a feature decoder, before sending the information to the next layer. With C3 block, each neuron accounts for the channel-wise responses from other neurons at the same layer and learns more discriminative and complementary representations. Extensive experiments for multiple computer vision tasks show that our proposed mechanism allows shallower networks to aggregate useful information within each layer, and performances outperform baseline deep networks and other competitive methods.

To date, top-performing optical flow estimation methods only take pairs of consecutive frames into account. While elegant and appealing, the idea of using more than two frames has not yet produced state-of-the-art results. We present a simple, yet effective fusion approach for multi-frame optical flow that benefits from longer-term temporal cues. Our method first warps the optical flow from previous frames to the current, thereby yielding multiple plausible estimates. It then fuses the complementary information carried by these estimates into a new optical flow field. At the time of writing, our method ranks first among published results in the MPI Sintel and KITTI 2015 benchmarks. Our models will be available on https://github.com/NVlabs/PWC-Net.

We develop a 3D object detection algorithm that uses latent support surfaces to capture contextual relationships in indoor scenes. Existing 3D representations for RGB-D images capture the local shape and appearance of object categories, but have limited power to represent objects with different visual styles. The detection of small objects is also challenging because the search space is very large in 3D scenes. However, we observe that much of the shape variation within 3D object categories can be explained by the location of a latent support surface, and smaller objects are often supported by larger objects. Therefore, we explicitly use latent support surfaces to better represent the 3D appearance of large objects, and provide contextual cues to improve the detection of small objects. We evaluate our model with 19 object categories from the SUN RGB-D database, and demonstrate state-of-the-art performance.

Given two consecutive frames from a pair of stereo cameras, 3D scene flow methods simultaneously estimate the 3D geometry and motion of the observed scene. Many existing approaches use superpixels for regularization, but may predict inconsistent shapes and motions inside rigidly moving objects. We instead assume that scenes consist of foreground objects rigidly moving in front of a static background, and use semantic cues to produce pixel-accurate scene flow estimates. Our cascaded classification framework accurately models 3D scenes by iteratively refining semantic segmentation masks, stereo correspondences, 3D rigid motion estimates, and optical flow fields. We evaluate our method on the challenging KITTI autonomous driving benchmark, and show that accounting for the motion of segmented vehicles leads to state-of-the-art performance.

We develop new representations and algorithms for three-dimensional (3D) object detection and spatial layout prediction in cluttered indoor scenes. RGB-D images are traditionally described by local geometric features of the 3D point cloud. We propose a cloud of oriented gradient (COG) descriptor that links the 2D appearance and 3D pose of object categories, and thus accurately models how perspective projection affects perceived image boundaries. We also propose a "Manhattan voxel" representation which better captures the 3D room layout geometry of common indoor environments. Effective classification rules are learned via a structured prediction framework that accounts for the intersection-over-union overlap of hypothesized 3D cuboids with human annotations, as well as orientation estimation errors. Contextual relationships among categories and layout are captured via a cascade of classifiers, leading to holistic scene hypotheses with improved accuracy. Our model is learned solely from annotated RGB-D images, without the benefit of CAD models, but nevertheless its performance substantially exceeds the state-of-the-art on the SUN RGB-D database. Avoiding CAD models allows easier learning of detectors for many object categories.

Recent developments in human-robot interaction brings about higher requirements for robot navigation. Existing Simultaneous Localization and Mapping (SLAM) algorithms face open challenges for navigation in complex dynamic environments due to presumptions of static environments or exceeding computational limitations. In this paper, we propose a robust graph SLAM formuation exploring Expectation Maximization algorithms to characterize landmark mobility while establishing the estimations of robot trajectory and the map. We evaluate the performance of existing robust SLAM algorithms as baselines, and validate the improvement of our new framework against datasets of dynamic environments with moving landmarks.

We live in a dynamic visual world where the appearance of scenes changes dramatically from hour to hour or season to season. In this work we study "transient scene attributes" -- high level properties which affect scene appearance, such as "snow", "autumn", "dusk", "fog". We define 40 transient attributes and use crowdsourcing to annotate thousands of images from 101 webcams. We use this "transient attribute database" to train regressors that can predict the presence of attributes in novel images. We demonstrate a photo organization method based on predicted attributes. Finally we propose a high-level image editing method which allows a user to adjust the attributes of a scene, e.g. change a scene to be "snowy" or "sunset". To support attribute manipulation we introduce a novel appearance transfer technique which is simple and fast yet competitive with the state-of-the-art. We show that we can convincingly modify many transient attributes in outdoor scenes.

We propose a hierarchical segmentation algorithm that starts with a very fine over segmentation and gradually merges regions using a cascade of boundary classifiers. This approach allows the weights of region and boundary features to adapt to the segmentation scale at which they are applied. The stages of the cascade are trained sequentially, with asymetric loss to maximize boundary recall. On six segmentation data sets, our algorithm achieves best performance under most region-quality measures, and does it with fewer segments than the prior work. Our algorithm is also highly competitive in a dense over segmentation (super pixel) regime under boundary-based measures.