Zihang Lai

Xuran Pan, Zihang Lai, Shiji Song et al.

Recently, Neural Radiance Fields (NeRF) has shown promising performances on reconstructing 3D scenes and synthesizing novel views from a sparse set of 2D images. Albeit effective, the performance of NeRF is highly influenced by the quality of training samples. With limited posed images from the scene, NeRF fails to generalize well to novel views and may collapse to trivial solutions in unobserved regions. This makes NeRF impractical under resource-constrained scenarios. In this paper, we present a novel learning framework, ActiveNeRF, aiming to model a 3D scene with a constrained input budget. Specifically, we first incorporate uncertainty estimation into a NeRF model, which ensures robustness under few observations and provides an interpretation of how NeRF understands the scene. On this basis, we propose to supplement the existing training set with newly captured samples based on an active learning scheme. By evaluating the reduction of uncertainty given new inputs, we select the samples that bring the most information gain. In this way, the quality of novel view synthesis can be improved with minimal additional resources. Extensive experiments validate the performance of our model on both realistic and synthetic scenes, especially with scarcer training data. Code will be released at \url{https://github.com/LeapLabTHU/ActiveNeRF}.

Yizeng Han, Yifan Pu, Zihang Lai et al.

Early exiting is an effective paradigm for improving the inference efficiency of deep networks. By constructing classifiers with varying resource demands (the exits), such networks allow easy samples to be output at early exits, removing the need for executing deeper layers. While existing works mainly focus on the architectural design of multi-exit networks, the training strategies for such models are largely left unexplored. The current state-of-the-art models treat all samples the same during training. However, the early-exiting behavior during testing has been ignored, leading to a gap between training and testing. In this paper, we propose to bridge this gap by sample weighting. Intuitively, easy samples, which generally exit early in the network during inference, should contribute more to training early classifiers. The training of hard samples (mostly exit from deeper layers), however, should be emphasized by the late classifiers. Our work proposes to adopt a weight prediction network to weight the loss of different training samples at each exit. This weight prediction network and the backbone model are jointly optimized under a meta-learning framework with a novel optimization objective. By bringing the adaptive behavior during inference into the training phase, we show that the proposed weighting mechanism consistently improves the trade-off between classification accuracy and inference efficiency. Code is available at https://github.com/LeapLabTHU/L2W-DEN.

Zhirong Wu, Zihang Lai, Xiao Sun et al.

The paper presents a scalable approach for learning spatially distributed visual representations over individual tokens and a holistic instance representation simultaneously. We use self-attention blocks to represent spatially distributed tokens, followed by cross-attention blocks to aggregate the holistic image instance. The core of the approach is the use of extremely large token masking (75\%-90\%) as the data augmentation for supervision. Our model, named ExtreMA, follows the plain BYOL approach where the instance representation from the unmasked subset is trained to predict that from the intact input. Instead of encouraging invariance across inputs, the model is required to capture informative variations in an image. The paper makes three contributions: 1) It presents random masking as a strong and computationally efficient data augmentation for siamese representation learning. 2) With multiple sampling per instance, extreme masking greatly speeds up learning and improves performance with more data. 3) ExtreMA obtains stronger linear probing performance than masked modeling methods, and better transfer performance than prior contrastive models.

Edgar Sucar, Eldar Insafutdinov, Zihang Lai et al.

Powerful 3D representations such as DUSt3R invariant point maps, which encode 3D shape and camera parameters, have significantly advanced feed forward 3D reconstruction. While point maps assume static scenes, Dynamic Point Maps (DPMs) extend this concept to dynamic 3D content by additionally representing scene motion. However, existing DPMs are limited to image pairs and, like DUSt3R, require post processing via optimization when more than two views are involved. We argue that DPMs are more useful when applied to videos and introduce V-DPM to demonstrate this. First, we show how to formulate DPMs for video input in a way that maximizes representational power, facilitates neural prediction, and enables reuse of pretrained models. Second, we implement these ideas on top of VGGT, a recent and powerful 3D reconstructor. Although VGGT was trained on static scenes, we show that a modest amount of synthetic data is sufficient to adapt it into an effective V-DPM predictor. Our approach achieves state of the art performance in 3D and 4D reconstruction for dynamic scenes. In particular, unlike recent dynamic extensions of VGGT such as P3, DPMs recover not only dynamic depth but also the full 3D motion of every point in the scene.

Zihang Lai, Eldar Insafutdinov, Edgar Sucar et al.

Dense point tracking is a fundamental problem in computer vision, with applications ranging from video analysis to robotic manipulation. State-of-the-art trackers typically rely on cost volumes to match features across frames, but this approach incurs quadratic complexity in spatial resolution, limiting scalability and efficiency. In this paper, we propose \method, a novel dense point tracker that eschews cost volumes in favor of warping. Inspired by recent advances in optical flow, our approach iteratively refines track estimates by warping features from the target frame to the query frame based on the current estimate. Combined with a transformer architecture that performs joint spatiotemporal reasoning across all tracks, our design establishes long-range correspondences without computing feature correlations. Our model is simple and achieves state-of-the-art performance on standard dense point tracking benchmarks, including TAP-Vid-DAVIS, TAP-Vid-Kinetics, and Robo-TAP. Remarkably, the model also excels at optical flow, sometimes outperforming specialized methods on the Sintel, KITTI, and Spring benchmarks. These results suggest that warping-based architectures can unify dense point tracking and optical flow estimation.

Yulin Wang, Yang Yue, Yuanze Lin et al.

Recent works have shown that the computational efficiency of video recognition can be significantly improved by reducing the spatial redundancy. As a representative work, the adaptive focus method (AdaFocus) has achieved a favorable trade-off between accuracy and inference speed by dynamically identifying and attending to the informative regions in each video frame. However, AdaFocus requires a complicated three-stage training pipeline (involving reinforcement learning), leading to slow convergence and is unfriendly to practitioners. This work reformulates the training of AdaFocus as a simple one-stage algorithm by introducing a differentiable interpolation-based patch selection operation, enabling efficient end-to-end optimization. We further present an improved training scheme to address the issues introduced by the one-stage formulation, including the lack of supervision, input diversity and training stability. Moreover, a conditional-exit technique is proposed to perform temporal adaptive computation on top of AdaFocus without additional training. Extensive experiments on six benchmark datasets (i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, and Jester) demonstrate that our model significantly outperforms the original AdaFocus and other competitive baselines, while being considerably more simple and efficient to train. Code is available at https://github.com/LeapLabTHU/AdaFocusV2.

Yan Wang, Zihang Lai, Gao Huang et al.

Many applications of stereo depth estimation in robotics require the generation of accurate disparity maps in real time under significant computational constraints. Current state-of-the-art algorithms force a choice between either generating accurate mappings at a slow pace, or quickly generating inaccurate ones, and additionally these methods typically require far too many parameters to be usable on power- or memory-constrained devices. Motivated by these shortcomings, we propose a novel approach for disparity prediction in the anytime setting. In contrast to prior work, our end-to-end learned approach can trade off computation and accuracy at inference time. Depth estimation is performed in stages, during which the model can be queried at any time to output its current best estimate. Our final model can process 1242$ \times $375 resolution images within a range of 10-35 FPS on an NVIDIA Jetson TX2 module with only marginal increases in error -- using two orders of magnitude fewer parameters than the most competitive baseline. The source code is available at https://github.com/mileyan/AnyNet .

Zeren Jiang, Yushi Lan, Yihang Luo et al.

Dense 3D reconstruction and tracking of dynamic scenes from monocular video remains an important open challenge in computer vision. Progress in this area has been constrained by the scarcity of high-quality datasets with dense, complete, and accurate geometric annotations. To address this limitation, we introduce Syn4D, a multiview synthetic dataset of dynamic scenes that includes ground-truth camera motion, depth maps, dense tracking, and parametric human pose annotations. A key feature of Syn4D is the ability to unproject any pixel into 3D to any time and to any camera. We conduct extensive evaluations across multiple downstream tasks to demonstrate the utility and effectiveness of the proposed dataset, including 4D scene reconstruction, 3D point tracking, geometry-aware camera retargeting, and human pose estimation. The experimental results highlight Syn4D's potential to facilitate research in dynamic scene understanding and spatiotemporal modeling.

Edgar Sucar, Zihang Lai, Eldar Insafutdinov et al.

DUSt3R has recently shown that one can reduce many tasks in multi-view geometry, including estimating camera intrinsics and extrinsics, reconstructing the scene in 3D, and establishing image correspondences, to the prediction of a pair of viewpoint-invariant point maps, i.e., pixel-aligned point clouds defined in a common reference frame. This formulation is elegant and powerful, but unable to tackle dynamic scenes. To address this challenge, we introduce the concept of Dynamic Point Maps (DPM), extending standard point maps to support 4D tasks such as motion segmentation, scene flow estimation, 3D object tracking, and 2D correspondence. Our key intuition is that, when time is introduced, there are several possible spatial and time references that can be used to define the point maps. We identify a minimal subset of such combinations that can be regressed by a network to solve the sub tasks mentioned above. We train a DPM predictor on a mixture of synthetic and real data and evaluate it across diverse benchmarks for video depth prediction, dynamic point cloud reconstruction, 3D scene flow and object pose tracking, achieving state-of-the-art performance. Code, models and additional results are available at https://www.robots.ox.ac.uk/~vgg/research/dynamic-point-maps/.

Zihang Lai

Open-vocabulary semantic segmentation models aim to accurately assign a semantic label to each pixel in an image from a set of arbitrary open-vocabulary texts. In order to learn such pixel-level alignment, current approaches typically rely on a combination of (i) image-level VL model (e.g. CLIP), (ii) ground truth masks, and (iii) custom grouping encoders. In this paper, we introduce S-Seg, a novel model that can achieve surprisingly strong performance without depending on any of the above elements. S-Seg leverages pseudo-mask and language to train a MaskFormer, and can be easily trained from publicly available image-text datasets. Contrary to prior works, our model directly trains for pixel-level features and language alignment. Once trained, S-Seg generalizes well to multiple testing datasets without requiring fine-tuning. In addition, S-Seg has the extra benefits of scalability with data and consistently improvement when augmented with self-training. We believe that our simple yet effective approach will serve as a solid baseline for future research.

Zihang Lai, Andrea Vedaldi

Temporal consistency is critical in video prediction to ensure that outputs are coherent and free of artifacts. Traditional methods, such as temporal attention and 3D convolution, may struggle with significant object motion and may not capture long-range temporal dependencies in dynamic scenes. To address this gap, we propose the Tracktention Layer, a novel architectural component that explicitly integrates motion information using point tracks, i.e., sequences of corresponding points across frames. By incorporating these motion cues, the Tracktention Layer enhances temporal alignment and effectively handles complex object motions, maintaining consistent feature representations over time. Our approach is computationally efficient and can be seamlessly integrated into existing models, such as Vision Transformers, with minimal modification. It can be used to upgrade image-only models to state-of-the-art video ones, sometimes outperforming models natively designed for video prediction. We demonstrate this on video depth prediction and video colorization, where models augmented with the Tracktention Layer exhibit significantly improved temporal consistency compared to baselines.

Chunjiang Ge, Rui Huang, Mixue Xie et al.

Unsupervised domain adaption (UDA) aims to adapt models learned from a well-annotated source domain to a target domain, where only unlabeled samples are given. Current UDA approaches learn domain-invariant features by aligning source and target feature spaces. Such alignments are imposed by constraints such as statistical discrepancy minimization or adversarial training. However, these constraints could lead to the distortion of semantic feature structures and loss of class discriminability. In this paper, we introduce a novel prompt learning paradigm for UDA, named Domain Adaptation via Prompt Learning (DAPL). In contrast to prior works, our approach makes use of pre-trained vision-language models and optimizes only very few parameters. The main idea is to embed domain information into prompts, a form of representations generated from natural language, which is then used to perform classification. This domain information is shared only by images from the same domain, thereby dynamically adapting the classifier according to each domain. By adopting this paradigm, we show that our model not only outperforms previous methods on several cross-domain benchmarks but also is very efficient to train and easy to implement.

Zihang Lai, Sifei Liu, Alexei A. Efros et al.

A video autoencoder is proposed for learning disentan- gled representations of 3D structure and camera pose from videos in a self-supervised manner. Relying on temporal continuity in videos, our work assumes that the 3D scene structure in nearby video frames remains static. Given a sequence of video frames as input, the video autoencoder extracts a disentangled representation of the scene includ- ing: (i) a temporally-consistent deep voxel feature to represent the 3D structure and (ii) a 3D trajectory of camera pose for each frame. These two representations will then be re-entangled for rendering the input video frames. This video autoencoder can be trained directly using a pixel reconstruction loss, without any ground truth 3D or camera pose annotations. The disentangled representation can be applied to a range of tasks, including novel view synthesis, camera pose estimation, and video generation by motion following. We evaluate our method on several large- scale natural video datasets, and show generalization results on out-of-domain images.

Zihang Lai, Senthil Purushwalkam, Abhinav Gupta

The ability to find correspondences in visual data is the essence of most computer vision tasks. But what are the right correspondences? The task of visual correspondence is well defined for two different images of same object instance. In case of two images of objects belonging to same category, visual correspondence is reasonably well-defined in most cases. But what about correspondence between two objects of completely different category -- e.g., a shoe and a bottle? Does there exist any correspondence? Inspired by humans' ability to: (a) generalize beyond semantic categories and; (b) infer functional affordances, we introduce the problem of functional correspondences in this paper. Given images of two objects, we ask a simple question: what is the set of correspondences between these two images for a given task? For example, what are the correspondences between a bottle and shoe for the task of pounding or the task of pouring. We introduce a new dataset: FunKPoint that has ground truth correspondences for 10 tasks and 20 object categories. We also introduce a modular task-driven representation for attacking this problem and demonstrate that our learned representation is effective for this task. But most importantly, because our supervision signal is not bound by semantics, we show that our learned representation can generalize better on few-shot classification problem. We hope this paper will inspire our community to think beyond semantics and focus more on cross-category generalization and learning representations for robotics tasks.

Zihang Lai, Erika Lu, Weidi Xie

Recent interest in self-supervised dense tracking has yielded rapid progress, but performance still remains far from supervised methods. We propose a dense tracking model trained on videos without any annotations that surpasses previous self-supervised methods on existing benchmarks by a significant margin (+15%), and achieves performance comparable to supervised methods. In this paper, we first reassess the traditional choices used for self-supervised training and reconstruction loss by conducting thorough experiments that finally elucidate the optimal choices. Second, we further improve on existing methods by augmenting our architecture with a crucial memory component. Third, we benchmark on large-scale semi-supervised video object segmentation(aka. dense tracking), and propose a new metric: generalizability. Our first two contributions yield a self-supervised network that for the first time is competitive with supervised methods on standard evaluation metrics of dense tracking. When measuring generalizability, we show self-supervised approaches are actually superior to the majority of supervised methods. We believe this new generalizability metric can better capture the real-world use-cases for dense tracking, and will spur new interest in this research direction.

Zihang Lai, Weidi Xie

The objective of this paper is self-supervised learning of feature embeddings that are suitable for matching correspondences along the videos, which we term correspondence flow. By leveraging the natural spatial-temporal coherence in videos, we propose to train a ``pointer'' that reconstructs a target frame by copying pixels from a reference frame. We make the following contributions: First, we introduce a simple information bottleneck that forces the model to learn robust features for correspondence matching, and prevent it from learning trivial solutions, \eg matching based on low-level colour information. Second, to tackle the challenges from tracker drifting, due to complex object deformations, illumination changes and occlusions, we propose to train a recursive model over long temporal windows with scheduled sampling and cycle consistency. Third, we achieve state-of-the-art performance on DAVIS 2017 video segmentation and JHMDB keypoint tracking tasks, outperforming all previous self-supervised learning approaches by a significant margin. Fourth, in order to shed light on the potential of self-supervised learning on the task of video correspondence flow, we probe the upper bound by training on additional data, \ie more diverse videos, further demonstrating significant improvements on video segmentation.

Zhihua Wang, Stefano Rosa, Yishu Miao et al.

Humans are able to make rich predictions about the future dynamics of physical objects from a glance. On the other hand, most existing computer vision approaches require strong assumptions about the underlying system, ad-hoc modeling, or annotated datasets, to carry out even simple predictions. To tackle this gap, we propose a new perspective on the problem of learning intuitive physics that is inspired by the spatial memory representation of objects and spaces in human brains, in particular the co-existence of egocentric and allocentric spatial representations. We present a generic framework that learns a layered representation of the physical world, using a cascade of invertible modules. In this framework, real images are first converted to a synthetic domain representation that reduces complexity arising from lighting and texture. Then, an allocentric viewpoint transformer removes viewpoint complexity by projecting images to a canonical view. Finally, a novel Recurrent Latent Variation Network (RLVN) architecture learns the dynamics of the objects interacting with the environment and predicts future motion, leveraging the availability of unlimited synthetic simulations. Predicted frames are then projected back to the original camera view and translated back to the real world domain. Experimental results show the ability of the framework to consistently and accurately predict several frames in the future and the ability to adapt to real images.

Junwen Bai, Zihang Lai, Runzhe Yang et al.

Many real-world tasks involve identifying patterns from data satisfying background or prior knowledge. In domains like materials discovery, due to the flaws and biases in raw experimental data, the identification of X-ray diffraction patterns (XRD) often requires a huge amount of manual work in finding refined phases that are similar to the ideal theoretical ones. Automatically refining the raw XRDs utilizing the simulated theoretical data is thus desirable. We propose imitation refinement, a novel approach to refine imperfect input patterns, guided by a pre-trained classifier incorporating prior knowledge from simulated theoretical data, such that the refined patterns imitate the ideal data. The classifier is trained on the ideal simulated data to classify patterns and learns an embedding space where each class is represented by a prototype. The refiner learns to refine the imperfect patterns with small modifications, such that their embeddings are closer to the corresponding prototypes. We show that the refiner can be trained in both supervised and unsupervised fashions. We further illustrate the effectiveness of the proposed approach both qualitatively and quantitatively in a digit refinement task and an X-ray diffraction pattern refinement task in materials discovery.