ICCV 2023 Papers

Continual Learning (CL) involves training a machine learning model in a sequential manner to learn new information while retaining previously learned tasks without the presence of previous training data. Although there has been significant interest in CL, most recent CL approaches in computer vision have focused on convolutional architectures only. However, with the recent success of vision transformers, there is a need to explore their potential for CL. Although there have been some recent CL approaches for vision transformers, they either store training instances of previous tasks or require a task identifier during test time, which can be limiting. This paper proposes a new exemplar-free approach for class/task incremental learning called ConTraCon, which does not require task-id to be explicitly present during inference and avoids the need for storing previous training instances. The proposed approach leverages the transformer architecture and involves re-weighting the key, query, and value weights of the multi-head self-attention layers of a transformer trained on a similar task. The re-weighting is done using convolution, which enables the approach to maintain low parameter requirements per task. Additionally, an image augmentation-based entropic task identification approach is used to predict tasks without requiring task-ids during inference. Experiments on four benchmark datasets demonstrate that the proposed approach outperforms several competitive approaches while requiring fewer parameters.

Autoexposure (AE) is a critical step applied by camera systems to ensure properly exposed images. While current AE algorithms are effective in well-lit environments with constant illumination, these algorithms still struggle in environments with bright light sources or scenes with abrupt changes in lighting. A significant hurdle in developing new AE algorithms for challenging environments, especially those with time-varying lighting, is the lack of suitable image datasets. To address this issue, we have captured a new 4D exposure dataset that provides a large solution space (i.e., shutter speed range from 1/500 to 15 seconds) over a temporal sequence with moving objects, bright lights, and varying lighting. In addition, we have designed a software platform to allow AE algorithms to be used in a plug-and-play manner with the dataset. Our dataset and associate platform enable repeatable evaluation of different AE algorithms and provide a much-needed starting point to develop better AE methods. We examine several existing AE strategies using our dataset and show that most users prefer a simple saliency method for challenging lighting conditions.

We present ExBluRF, a novel view synthesis method for extreme motion blurred images based on efficient radiance fields optimization. Our approach consists of two main components: 6-DOF camera trajectory-based motion blur formulation and voxel-based radiance fields. From extremely blurred images, we optimize the sharp radiance fields by jointly estimating the camera trajectories that generate the blurry images. In training, multiple rays along the camera trajectory are accumulated to reconstruct single blurry color, which is equivalent to the physical motion blur operation. We minimize the photo-consistency loss on blurred image space and obtain the sharp radiance fields with camera trajectories that explain the blur of all images. The joint optimization on the blurred image space demands painfully increasing computation and resources proportional to the blur size. Our method solves this problem by replacing the MLP-based framework to low-dimensional 6-DOF camera poses and voxel-based radiance fields. Compared with the existing works, our approach restores much sharper 3D scenes from challenging motion blurred views with the order of 10x less training time and GPU memory consumption.

Because of the diversity in lighting environments, existing illumination estimation techniques have been designed explicitly on indoor or outdoor environments. Methods have focused specifically on capturing accurate energy (e.g., through parametric lighting models), which emphasizes shading and strong cast shadows; or producing plausible texture (e.g., with GANs), which prioritizes plausible reflections. Approaches which provide editable lighting capabilities have been proposed, but these tend to be with simplified lighting models, offering limited realism. In this work, we propose to bridge the gap between these recent trends in the literature, and propose a method which combines a parametric light model with 360deg panoramas, ready to use as HDRI in rendering engines. We leverage recent advances in GAN-based LDR panorama extrapolation from a regular image, which we extend to HDR using parametric spherical gaussians. To achieve this, we introduce a novel lighting co-modulation method that injects lighting-related features throughout the generator, tightly coupling the original or edited scene illumination within the panorama generation process. In our representation, users can easily edit light direction, intensity, number, etc. to impact shading while providing rich, complex reflections while seamlessly blending with the edits. Furthermore, our method encompasses indoor and outdoor environments, demonstrating state-of-the-art results even when compared to domain-specific methods.

Vision Transformers achieve impressive accuracy across a range of visual recognition tasks. Unfortunately, their accuracy frequently comes with high computational costs. This is a particular issue in video recognition, where models are often applied repeatedly across frames or temporal chunks. In this work, we exploit temporal redundancy between subsequent inputs to reduce the cost of Transformers for video processing. We describe a method for identifying and re-processing only those tokens that have changed significantly over time. Our proposed family of models, Eventful Transformers, can be converted from existing Transformers (often without any re-training) and give adaptive control over the compute cost at runtime. We evaluate our method on large-scale datasets for video object detection (ImageNet VID) and action recognition (EPIC-Kitchens 100). Our approach leads to significant computational savings (on the order of 2-4x) with only minor reductions in accuracy.

Event cameras provide an advantage over traditional frame-based cameras when capturing fast-moving objects without a motion blur. They achieve this by recording changes in light intensity (known as events), thus allowing them to operate at a much higher frequency and making them suitable for capturing motions in a highly dynamic scene. Many recent studies have proposed methods to train neural networks (NNs) for predicting optical flow from events. However, they often rely on a spatio-temporal representation constructed from events over a fixed interval, such as 10Hz used in training on the DSEC dataset. This limitation restricts the flow prediction to the same interval (10Hz) whereas the fast speed of event cameras, which can operate up to 3kHz, has not been effectively utilized. In this work, we show that a temporally dense flow estimation at 100Hz can be achieved by treating the flow estimation as a sequential problem using two different variants of recurrent networks - Long-short term memory (LSTM) and spiking neural network (SNN). First, We utilize the NN model constructed similar to the popular EV-FlowNet but with LSTM layers to demonstrate the efficiency of our training method. The model not only produces 10x more frequent optical flow than the existing ones, but the estimated flows also have 13% lower errors than predictions from the baseline EV-FlowNet. Second, we construct an EV-FlowNet SNN but with leaky integrate and fire neurons to efficiently capture the temporal dynamics. We found that simple inherent recurrent dynamics of SNN lead to significant parameter reduction compared to the LSTM model. In addition, because of its event-driven computation, the spiking model is estimated to consume only 1.5% energy of the LSTM model, highlighting the efficiency of SNN in processing events and the potential for achieving temporally dense flow.

In this work, we focus on the task of procedure planning from instructional videos with text supervision, where a model aims to predict an action sequence to transform the initial visual state into the goal visual state. A critical challenge of this task is the large semantic gap between observed visual states and unobserved intermediate actions, which is ignored by previous works. Specifically, this semantic gap refers to that the contents in the observed visual states are semantically different from the elements of some action text labels in a procedure. To bridge this semantic gap, we propose a novel event-guided paradigm, which first infers events from the observed states and then plans out actions based on both the states and predicted events. Our inspiration comes from that planning a procedure from an instructional video is to complete a specific event and a specific event usually involves specific actions. Based on the proposed paradigm, we contribute an Event-guided Prompting-based Procedure Planning (E3P) model, which encodes event information into the sequential modeling process to support procedure planning. To further consider the strong action associations within each event, our E3P adopts a mask-and-predict approach for relation mining, incorporating a probabilistic masking scheme for regularization. Extensive experiments on three datasets demonstrate the effectiveness of our proposed model.

This paper proposes a pre-trained neural network for handling event camera data. Our model is a self-supervised learning framework, and uses paired event camera data and natural RGB images for training. Our method contains three modules connected in a sequence: i) a family of event data augmentations, generating meaningful event images for self-supervised training; ii) a conditional masking strategy to sample informative event patches from event images, encouraging our model to capture the spatial layout of a scene and accelerating training; iii) a contrastive learning approach, enforcing the similarity of embeddings between matching event images, and between paired event and RGB images. An embedding projection loss is proposed to avoid the model collapse when enforcing the event image embedding similarities. A probability distribution alignment loss is proposed to encourage the event image to be consistent with its paired RGB image in the feature space. Transfer learning performance on downstream tasks shows the superiority of our method over state-of-the-art methods. For example, we achieve top-1 accuracy at 64.83% on the N-ImageNet dataset. Our code is available at https://github.com/Yan98/Event-Camera-Data-Pre-training.

Part-prototype networks (e.g., ProtoPNet, ProtoTree, and ProtoPool) have attracted broad research interest for their intrinsic interpretability and comparable accuracy to non-interpretable counterparts. However, recent works find that the interpretability from prototypes is fragile, due to the semantic gap between the similarities in the feature space and that in the input space. In this work, we strive to address this challenge by making the first attempt to quantitatively and objectively evaluate the interpretability of the part-prototype networks. Specifically, we propose two evaluation metrics, termed as "consistency score" and "stability score", to evaluate the explanation consistency across images and the explanation robustness against perturbations, respectively, both of which are essential for explanations taken into practice. Furthermore, we propose an elaborated part-prototype network with a shallow-deep feature alignment (SDFA) module and a score aggregation (SA) module to improve the interpretability of prototypes. We conduct systematical evaluation experiments and provide substantial discussions to uncover the interpretability of existing part-prototype networks. Experiments on three benchmarks across nine architectures demonstrate that our model achieves significantly superior performance to the state of the art, in both the accuracy and interpretability. Our code is available at https://github.com/hqhQAQ/EvalProtoPNet.

While large text-to-image models are able to synthesize "novel" images, these images are necessarily a reflection of the training data. The problem of data attribution in such models -- which of the images in the training set are most responsible for the appearance of a given generated image -- is a difficult yet important one. As an initial step toward this problem, we evaluate attribution through "customization" methods, which tune an existing large-scale model toward a given exemplar object or style. Our key insight is that this allows us to efficiently create synthetic images that are computationally influenced by the exemplar by construction. With our new dataset of such exemplar-influenced images, we are able to evaluate various data attribution algorithms and different possible feature spaces. Furthermore, by training on our dataset, we can tune standard models, such as DINO, CLIP, and ViT, toward the attribution problem. Even though the procedure is tuned towards small exemplar sets, we show generalization to larger sets. Finally, by taking into account the inherent uncertainty of the problem, we can assign soft attribution scores over a set of training images.

Single-photon sensors measure light signals at the finest possible resolution -- individual photons. These sensors introduce two major challenges in the form of strong Poisson noise and extremely large data acquisition rates, which are also inherited by downstream computer vision tasks. Previous work has largely focused on solving the image reconstruction problem first and then using off-the-shelf methods for downstream tasks, but the most general solutions that account for motion are costly and not scalable to large data volumes produced by single-photon sensors. This work forgoes the image reconstruction problem. Instead, we demonstrate computationally light-weight phase-based algorithms for the tasks of edge detection and motion estimation. These methods directly process the raw single-photon data as a 3D volume with a bank of velocity-tuned filters, achieving speed-ups of more than two orders of magnitude compared to explicit reconstruction-based methods. Project webpage: https://wisionlab.com/project/eulerian-single-photon-vision/

Binarization of neural networks is a dominant paradigm in neural networks compression. The pioneering work BinaryConnect uses Straight Through Estimator (STE) to mimic the gradients of the sign function, but it also causes the crucial inconsistency problem. Most of the previous methods design different estimators instead of STE to mitigate it. However, they ignore the fact that when reducing the estimating error, the gradient stability will decrease concomitantly. These highly divergent gradients will harm the model training and increase the risk of gradient vanishing and gradient exploding. To fully take the gradient stability into consideration, we present a new perspective to the BNNs training, regarding it as the equilibrium between the estimating error and the gradient stability. In this view, we firstly design two indicators to quantitatively demonstrate the equilibrium phenomenon. In addition, in order to balance the estimating error and the gradient stability well, we revise the original straight through estimator and propose a power function based estimator, Rectified Straight Through Estimator (ReSTE for short). Comparing to other estimators, ReSTE is rational and capable of flexibly balancing the estimating error with the gradient stability. Extensive experiments on CIFAR-10 and ImageNet datasets show that ReSTE has excellent performance and surpasses the state-of-the-art methods without any auxiliary modules or losses.

We introduce a novel method to estimate the essential matrix for two-view Structure from Motion (SfM). We show that every 3 by 3 essential matrix can be embedded in a 4 by 4 rotation, having its bottom right entry fixed to zero; we call the latter the quintessential matrix. This embedding leads to rich relations with the space of 4-D rotations, quaternions, and the classical twisted-pair ambiguity in two-view SfM. We use this structure to derive a succession of semidefinite relaxations that require fewer parameters than the existing non-minimal solvers and yield faster convergence with certifiable optimality. We then exploit the low-rank geometry of these relaxations to reduce them to an equivalent optimization on a Riemannian manifold and solve them via the Riemannian Staircase method. The experimental evaluation confirms that our algorithm always finds the globally optimal solution and outperforms the existing non-minimal methods. We make our implementations open source.

Motivated by concerns that large-scale diffusion models can produce undesirable output such as sexually explicit content or copyrighted artistic styles, we study erasure of specific concepts from diffusion model weights. We propose a fine-tuning method that can erase a visual concept from a pre-trained diffusion model, given only the name of the style and using negative guidance as a teacher. We benchmark our method against previous approaches that remove sexually explicit content and demonstrate its effectiveness, performing on par with Safe Latent Diffusion and censored training. To evaluate artistic style removal, we conduct experiments erasing five modern artists from the network and conduct a user study to assess the human perception of the removed styles. Unlike previous methods, our approach can remove concepts from a diffusion model permanently rather than modifying the output at the inference time, so it cannot be circumvented even if a user has access to model weights. Our code, data, and results are available at erasing.baulab.info

The scene graph generation (SGG) task is designed to identify the predicates based on the subject-object pairs. However, existing datasets generally include two imbalance cases: one is the class imbalance from the predicted predicates and another is the context imbalance from the given subject-object pairs, which presents significant challenges for SGG. Most existing methods focus on the imbalance of the predicted predicate while ignoring the imbalance of the subject-object pairs, which could not achieve satisfactory results. To address the two imbalance cases, we propose a novel Environment Invariant Curriculum Relation learning (EICR) method, which can be applied in a plug-and-play fashion to existing SGG methods. Concretely, to remove the imbalance of the subject-object pairs, we first construct different distribution environments for the subject-object pairs and learn a model invariant to the environment changes. Then, we construct a class-balanced curriculum learning strategy to balance the different environments to remove the predicate imbalance. Comprehensive experiments conducted on VG and GQA datasets demonstrate that our EICR framework can be taken as a general strategy for various SGG models, and achieve significant improvements.

Generalization capability of vision-based deep reinforcement learning (RL) is indispensable to deal with dynamic environment changes that exist in visual observations. The high-dimensional space of the visual input, however, imposes challenges in adapting an agent to unseen environments. In this work, we propose Environment Agnostic Reinforcement learning (EAR), which is a compact framework for domain generalization of the visual deep RL. Environment-agnostic features (EAFs) are extracted by leveraging three novel objectives based on feature factorization, reconstruction, and episode-aware state shifting, so that policy learning is accomplished only with vital features. EAR is a simple single-stage method with a low model complexity and a fast inference time, ensuring a high reproducibility, while attaining state-of-the-art performance in the DeepMind Control Suite and DrawerWorld benchmarks.

Federated learning (FL) is a privacy-enhanced distributed machine learning framework, in which multiple clients collaboratively train a global model by exchanging their model updates without sharing local private data. However, the adversary can use gradient inversion attacks to reveal the clients' privacy from the shared model updates. Previous attacks assume the adversary can infer the local learning rate of each client, while we observe that: (1) using the uniformly distributed random local learning rates does not incur much accuracy loss of the global model, and (2) personalizing local learning rates can mitigate the drift issue which is caused by non-IID (identically and independently distributed) data. Moreover, we theoretically derive a convergence guarantee to FedAvg with uniformly perturbed local learning rates. Therefore, by perturbing the learning rate of each client with random noise, we propose a learning rate perturbation (LRP) defense against gradient inversion attacks. Specifically, for classification tasks, we adapt LPR to ada-LPR by personalizing the expectation of each local learning rate. The experiments show that our defenses can well enhance privacy preservation against existing gradient inversion attacks, and LRP outperforms 5 baseline defenses against a state-of-the-art gradient inversion attack. In addition, our defenses only incur minor accuracy reductions (less than 0.5%) of the global model. So they are effective in real applications.

Recovering information from non-line-of-sight (NLOS) imaging is a computationally-intensive inverse problem. Most physics-based NLOS imaging methods address the complexity of this problem by assuming three-bounce reflections and no self-occlusion. However, these assumptions may break down for objects with large depth variations, preventing physics-based algorithms from accurately reconstructing the details and high-frequency information. On the other hand, while learning-based methods can avoid these assumptions, they may struggle to reconstruct details without specific designs due to the spectral bias of neural networks. To overcome these issues, we propose a novel approach that enhances physics-based NLOS imaging methods by introducing a learnable inverse kernel in the Fourier domain and using an attention mechanism to improve the neural network to learn high-frequency information. Our method is evaluated on publicly available and new synthetic datasets, demonstrating its commendable performance compared to prior physics-based and learning-based methods, especially for objects with large depth variations. Moreover, our approach generalizes well to real data and can be applied to tasks such as classification and depth reconstruction. We will make our code and dataset publicly available: https://sci2020.github.io.

Cross-scene generalizable NeRF models, which can directly synthesize novel views of unseen scenes, have become a new spotlight of the NeRF field. Several existing attempts rely on increasingly end-to-end "neuralized" architectures, i.e., replacing scene representation and/or rendering modules with performant neural networks such as transformers, and turning novel view synthesis into a feed-forward inference pipeline. While those feedforward "neuralized" architectures still do not fit diverse scenes well out of the box, we propose to bridge them with the powerful Mixture-of-Experts (MoE) idea from large language models (LLMs), which has demonstrated superior generalization ability by balancing between larger overall model capacity and flexible per-instance specialization. Starting from a recent generalizable NeRF architecture called GNT, we first demonstrate that MoE can be neatly plugged in to enhance the model. We further customize a shared permanent expert and a geometry-aware consistency loss to enforce cross-scene consistency and spatial smoothness respectively, which are essential for generalizable view synthesis. Our proposed model, dubbed GNT with Mixture-of-View-Experts (GNT-MOVE), has experimentally shown state-of-the-art results when transferring to unseen scenes, indicating remarkably better cross-scene generalization in both zero-shot and few-shot settings. Our codes are available at https://github.com/VITA-Group/GNT-MOVE.