CVPR 2023 Papers

Confidence-based pseudo-labeling is among the dominant approaches in semi-supervised learning (SSL). It relies on including high-confidence predictions made on unlabeled data as additional targets to train the model. We propose ProtoCon, a novel SSL method aimed at the less-explored label-scarce SSL where such methods usually underperform. ProtoCon refines the pseudo-labels by leveraging their nearest neighbours' information. The neighbours are identified as the training proceeds using an online clustering approach operating in an embedding space trained via a prototypical loss to encourage well-formed clusters. The online nature of ProtoCon allows it to utilise the label history of the entire dataset in one training cycle to refine labels in the following cycle without the need to store image embeddings. Hence, it can seamlessly scale to larger datasets at a low cost. Finally, ProtoCon addresses the poor training signal in the initial phase of training (due to fewer confident predictions) by introducing an auxiliary self-supervised loss. It delivers significant gains and faster convergence over state-of-the-art across 5 datasets, including CIFARs, ImageNet and DomainNet.

Motion forecasting is a key module in an autonomous driving system. Due to the heterogeneous nature of multi-sourced input, multimodality in agent behavior, and low latency required by onboard deployment, this task is notoriously challenging. To cope with these difficulties, this paper proposes a novel agent-centric model with anchor-informed proposals for efficient multimodal motion forecasting. We design a modality-agnostic strategy to concisely encode the complex input in a unified manner. We generate diverse proposals, fused with anchors bearing goal-oriented context, to induce multimodal prediction that covers a wide range of future trajectories. The network architecture is highly uniform and succinct, leading to an efficient model amenable for real-world deployment. Experiments reveal that our agent-centric network compares favorably with the state-of-the-art methods in prediction accuracy, while achieving scene-centric level inference latency.

In this paper, we propose an efficient event-based motion estimation framework for various motion models. Different from previous works, we design a progressive event-to-map alignment scheme and utilize the spatio-temporal correlations to align events. In detail, we progressively align sampled events in an event batch to the time-surface map and obtain the updated motion model by minimizing a novel time-surface loss. In addition, a dynamic batch size strategy is applied to adaptively adjust the batch size so that all events in the batch are consistent with the current motion model. Our framework has three advantages: a) the progressive scheme refines motion parameters iteratively, achieving accurate motion estimation; b) within one iteration, only a small portion of events are involved in optimization, which greatly reduces the total runtime; c) the dynamic batch size strategy ensures that the constant velocity assumption always holds. We conduct comprehensive experiments to evaluate our framework on challenging high-speed scenes with three motion models: rotational, homography, and 6-DOF models. Experimental results demonstrate that our framework achieves state-of-the-art estimation accuracy and efficiency.

Recently, transformer-based methods have gained significant success in sequential 2D-to-3D lifting human pose estimation. As a pioneering work, PoseFormer captures spatial relations of human joints in each video frame and human dynamics across frames with cascaded transformer layers and has achieved impressive performance. However, in real scenarios, the performance of PoseFormer and its follow-ups is limited by two factors: (a) The length of the input joint sequence; (b) The quality of 2D joint detection. Existing methods typically apply self-attention to all frames of the input sequence, causing a huge computational burden when the frame number is increased to obtain advanced estimation accuracy, and they are not robust to noise naturally brought by the limited capability of 2D joint detectors. In this paper, we propose PoseFormerV2, which exploits a compact representation of lengthy skeleton sequences in the frequency domain to efficiently scale up the receptive field and boost robustness to noisy 2D joint detection. With minimum modifications to PoseFormer, the proposed method effectively fuses features both in the time domain and frequency domain, enjoying a better speed-accuracy trade-off than its precursor. Extensive experiments on two benchmark datasets (i.e., Human3.6M and MPI-INF-3DHP) demonstrate that the proposed approach significantly outperforms the original PoseFormer and other transformer-based variants. Code is released at https://github.com/QitaoZhao/PoseFormerV2.

The remarkable breakthroughs in point cloud representation learning have boosted their usage in real-world applications such as self-driving cars and virtual reality. However, these applications usually have an urgent requirement for not only accurate but also efficient 3D object detection. Recently, knowledge distillation has been proposed as an effective model compression technique, which transfers the knowledge from an over-parameterized teacher to a lightweight student and achieves consistent effectiveness in 2D vision. However, due to point clouds' sparsity and irregularity, directly applying previous image-based knowledge distillation methods to point cloud detectors usually leads to unsatisfactory performance. To fill the gap, this paper proposes PointDistiller, a structured knowledge distillation framework for point clouds-based 3D detection. Concretely, PointDistiller includes local distillation which extracts and distills the local geometric structure of point clouds with dynamic graph convolution and reweighted learning strategy, which highlights student learning on the critical points or voxels to improve knowledge distillation efficiency. Extensive experiments on both voxels-based and raw points-based detectors have demonstrated the effectiveness of our method over seven previous knowledge distillation methods. For instance, our 4X compressed PointPillars student achieves 2.8 and 3.4 mAP improvements on BEV and 3D object detection, outperforming its teacher by 0.9 and 1.8 mAP, respectively. Codes are available in the supplementary material and will be released on Github.

Thanks to advances in computer vision and AI, there has been a large growth in the demand for cloud-based visual analytics in which images captured by a low-powered edge device are transmitted to the cloud for analytics. Use of conventional codecs (JPEG, MPEG, HEVC, etc.) for compressing such data introduces artifacts that can seriously degrade the performance of the downstream analytic tasks. Split-DNN computing has emerged as a paradigm to address such usages, in which a DNN is partitioned into a client-side portion and a server side portion. Low-complexity neural networks called 'bottleneck units' are introduced at the split point to transform the intermediate layer features into a lower-dimensional representation better suited for compression and transmission. Optimizing the pipeline for both compression and task-performance requires high-quality estimates of the information-theoretic rate of the intermediate features. Most works on compression for image analytics use heuristic approaches to estimate the rate, leading to suboptimal performance. We propose a high-quality 'neural rate-estimator' to address this gap. We interpret the lower-dimensional bottleneck output as a latent representation of the intermediate feature and cast the rate-distortion optimization problem as one of training an equivalent variational auto-encoder with an appropriate loss function. We show that this leads to improved rate-distortion outcomes. We further show that replacing supervised loss terms (such as cross-entropy loss) by distillation-based losses in a teacher-student framework allows for unsupervised training of bottleneck units without the need for explicit training labels. This makes our method very attractive for real world deployments where access to labeled training data is difficult or expensive. We demonstrate that our method outperforms several state-of-the-art methods by obtaining improved task accuracy at lower bitrates on image classification and semantic segmentation tasks.

We address efficient and structure-aware 3D scene representation from images. Nerflets are our key contribution-- a set of local neural radiance fields that together represent a scene. Each nerflet maintains its own spatial position, orientation, and extent, within which it contributes to panoptic, density, and radiance reconstructions. By leveraging only photometric and inferred panoptic image supervision, we can directly and jointly optimize the parameters of a set of nerflets so as to form a decomposed representation of the scene, where each object instance is represented by a group of nerflets. During experiments with indoor and outdoor environments, we find that nerflets: (1) fit and approximate the scene more efficiently than traditional global NeRFs, (2) allow the extraction of panoptic and photometric renderings from arbitrary views, and (3) enable tasks rare for NeRFs, such as 3D panoptic segmentation and interactive editing.

While some studies have proven that Swin Transformer (Swin) with window self-attention (WSA) is suitable for single image super-resolution (SR), the plain WSA ignores the broad regions when reconstructing high-resolution images due to a limited receptive field. In addition, many deep learning SR methods suffer from intensive computations. To address these problems, we introduce the N-Gram context to the low-level vision with Transformers for the first time. We define N-Gram as neighboring local windows in Swin, which differs from text analysis that views N-Gram as consecutive characters or words. N-Grams interact with each other by sliding-WSA, expanding the regions seen to restore degraded pixels. Using the N-Gram context, we propose NGswin, an efficient SR network with SCDP bottleneck taking multi-scale outputs of the hierarchical encoder. Experimental results show that NGswin achieves competitive performance while maintaining an efficient structure when compared with previous leading methods. Moreover, we also improve other Swin-based SR methods with the N-Gram context, thereby building an enhanced model: SwinIR-NG. Our improved SwinIR-NG outperforms the current best lightweight SR approaches and establishes state-of-the-art results. Codes are available at https://github.com/rami0205/NGramSwin.

Model-based deep learning has achieved astounding successes due in part to the availability of large-scale real-world data. However, processing such massive amounts of data comes at a considerable cost in terms of computations, storage, training and the search for good neural architectures. Dataset distillation has thus recently come to the fore. This paradigm involves distilling information from large real-world datasets into tiny and compact synthetic datasets such that processing the latter yields similar performances as the former. State-of-the-art methods primarily rely on learning the synthetic dataset by matching the gradients obtained during training between the real and synthetic data. However, these gradient-matching methods suffer from the accumulated trajectory error caused by the discrepancy between the distillation and subsequent evaluation. To alleviate the adverse impact of this accumulated trajectory error, we propose a novel approach that encourages the optimization algorithm to seek a flat trajectory. We show that the weights trained on synthetic data are robust against the accumulated errors perturbations with the regularization towards the flat trajectory. Our method, called Flat Trajectory Distillation (FTD), is shown to boost the performance of gradient-matching methods by up to 4.7% on a subset of images of the ImageNet dataset with higher resolution images. We also validate the effectiveness and generalizability of our method with datasets of different resolutions and demonstrate its applicability to neural architecture search.

This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps. Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves 84.0% ImageNet top-1 accuracy, outperforming the baseline of directly distilling a fine-tuned ViT-L by 1.2%. More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with 95% masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of 83.6%; surprisingly, it can still secure 82.4% top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code will be made publicly available.

Autonomous driving systems rely heavily on the underlying perception module which needs to be both performant and efficient to allow precise decisions in real-time. Avoiding collisions with pedestrians is of topmost priority in any autonomous driving system. Therefore, pedestrian detection is one of the core parts of such systems' perception modules. Current state-of-the-art pedestrian detectors have two major issues. Firstly, they have long inference times which affect the efficiency of the whole perception module, and secondly, their performance in the case of small and heavily occluded pedestrians is poor. We propose Localized Semantic Feature Mixers (LSFM), a novel, anchor-free pedestrian detection architecture. It uses our novel Super Pixel Pyramid Pooling module instead of the, computationally costly, Feature Pyramid Networks for feature encoding. Moreover, our MLPMixer-based Dense Focal Detection Network is used as a light detection head, reducing computational effort and inference time compared to existing approaches. To boost the performance of the proposed architecture, we adapt and use mixup augmentation which improves the performance, especially in small and heavily occluded cases. We benchmark LSFM against the state-of-the-art on well-established traffic scene pedestrian datasets. The proposed LSFM achieves state-of-the-art performance in Caltech, City Persons, Euro City Persons, and TJU-Traffic-Pedestrian datasets while reducing the inference time on average by 55%. Further, LSFM beats the human baseline for the first time in the history of pedestrian detection. Finally, we conducted a cross-dataset evaluation which proved that our proposed LSFM generalizes well to unseen data.

Recent DEtection TRansformer-based (DETR) models have obtained remarkable performance. Its success cannot be achieved without the re-introduction of multi-scale feature fusion in the encoder. However, the excessively increased tokens in multi-scale features, especially for about 75% of low-level features, are quite computationally inefficient, which hinders real applications of DETR models. In this paper, we present Lite DETR, a simple yet efficient end-to-end object detection framework that can effectively reduce the GFLOPs of the detection head by 60% while keeping 99% of the original performance. Specifically, we design an efficient encoder block to update high-level features (corresponding to small-resolution feature maps) and low-level features (corresponding to large-resolution feature maps) in an interleaved way. In addition, to better fuse cross-scale features, we develop a key-aware deformable attention to predict more reliable attention weights. Comprehensive experiments validate the effectiveness and efficiency of the proposed Lite DETR, and the efficient encoder strategy can generalize well across existing DETR-based models. The code will be released after the blind review.

Image resampling is a basic technique that is widely employed in daily applications. Existing deep neural networks (DNNs) have made impressive progress in resampling performance. Yet these methods are still not the perfect substitute for interpolation, due to the issues of efficiency and continuous resampling. In this work, we propose a novel method of Learning Resampling Function (termed LeRF), which takes advantage of both the structural priors learned by DNNs and the locally continuous assumption of interpolation methods. Specifically, LeRF assigns spatially-varying steerable resampling functions to input image pixels and learns to predict the hyper-parameters that determine the orientations of these resampling functions with a neural network. To achieve highly efficient inference, we adopt look-up tables (LUTs) to accelerate the inference of the learned neural network. Furthermore, we design a directional ensemble strategy and edge-sensitive indexing patterns to better capture local structures. Extensive experiments show that our method runs as fast as interpolation, generalizes well to arbitrary transformations, and outperforms interpolation significantly, e.g., up to 3dB PSNR gain over bicubic for x2 upsampling on Manga109.

Recent advance in 2D CNNs has revealed that large kernels are important. However, when directly applying large convolutional kernels in 3D CNNs, severe difficulties are met, where those successful module designs in 2D become surprisingly ineffective on 3D networks, including the popular depth-wise convolution. To address this vital challenge, we instead propose the spatial-wise partition convolution and its large-kernel module. As a result, it avoids the optimization and efficiency issues of naive 3D large kernels. Our large-kernel 3D CNN network, LargeKernel3D, yields notable improvement in 3D tasks of semantic segmentation and object detection. It achieves 73.9% mIoU on the ScanNetv2 semantic segmentation and 72.8% NDS nuScenes object detection benchmarks, ranking 1st on the nuScenes LIDAR leaderboard. The performance further boosts to 74.2% NDS with a simple multi-modal fusion. In addition, LargeKernel3D can be scaled to 17x17x17 kernel size on Waymo 3D object detection. For the first time, we show that large kernels are feasible and essential for 3D visual tasks.

We present a simple yet effective self-supervised pretraining method for image harmonization which can leverage large-scale unannotated image datasets. To achieve this goal, we first generate pre-training data online with our Label-Efficient Masked Region Transform (LEMaRT) pipeline. Given an image, LEMaRT generates a foreground mask and then applies a set of transformations to perturb various visual attributes, e.g., defocus blur, contrast, saturation, of the region specified by the generated mask. We then pre-train image harmonization models by recovering the original image from the perturbed image. Secondly, we introduce an image harmonization model, namely SwinIH, by retrofitting the Swin Transformer [27] with a combination of local and global self-attention mechanisms. Pretraining SwinIH with LEMaRT results in a new state of the art for image harmonization, while being label-efficient, i.e., consuming less annotated data for fine-tuning than existing methods. Notably, on iHarmony4 dataset [8], SwinIH outperforms the state of the art, i.e., SCS-Co [16] by a margin of 0.4 dB when it is fine-tuned on only 50% of the training data, and by 1.0 dB when it is trained on the full training dataset.

Rolling shutter correction (RSC) is becoming increasingly popular for RS cameras that are widely used in commercial and industrial applications. Despite the promising performance, existing RSC methods typically employ a two-stage network structure that ignores intrinsic information interactions and hinders fast inference. In this paper, we propose a single-stage encoder-decoder-based network, named JAMNet, for efficient RSC. It first extracts pyramid features from consecutive RS inputs, and then simultaneously refines the two complementary information (i.e., global shutter appearance and undistortion motion field) to achieve mutual promotion in a joint learning decoder. To inject sufficient motion cues for guiding joint learning, we introduce a transformer-based motion embedding module and propose to pass hidden states across pyramid levels. Moreover, we present a new data augmentation strategy "vertical flip + inverse order" to release the potential of the RSC datasets. Experiments on various benchmarks show that our approach surpasses the state-of-the-art methods by a large margin, especially with a 4.7 dB PSNR leap on real-world RSC. Code is available at https://github.com/GitCVfb/JAMNet.

Recovery of an underlying scene geometry from multi-view images stands as a long-time challenge in computer vision research. The recent promise leverages neural implicit surface learning and differentiable volume rendering, and achieves both the recovery of scene geometry and synthesis of novel views, where deep priors of neural models are used as an inductive smoothness bias. While promising for object-level surfaces, these methods suffer when coping with complex scene surfaces. In the meanwhile, traditional multi-view stereo can recover the geometry of scenes with rich textures, by globally optimizing the local, pixel-wise correspondences across multiple views. We are thus motivated to make use of the complementary benefits from the two strategies, and propose a method termed Helix-shaped neural implicit Surface learning or HelixSurf; HelixSurf uses the intermediate prediction from one strategy as the guidance to regularize the learning of the other one, and conducts such intertwined regularization iteratively during the learning process. We also propose an efficient scheme for differentiable volume rendering in HelixSurf. Experiments on surface reconstruction of indoor scenes show that our method compares favorably with existing methods and is orders of magnitude faster, even when some of existing methods are assisted with auxiliary training data. The source code is available at https://github.com/Gorilla-Lab-SCUT/HelixSurf.