CVPR 2023 Papers

Data-free Knowledge Distillation (DFKD) has gained popularity recently, with the fundamental idea of carrying out knowledge transfer from a Teacher neural network to a Student neural network in the absence of training data. However, in the Adversarial DFKD framework, the student network's accuracy, suffers due to the non-stationary distribution of the pseudo-samples under multiple generator updates. To this end, at every generator update, we aim to maintain the student's performance on previously encountered examples while acquiring knowledge from samples of the current distribution. Thus, we propose a meta-learning inspired framework by treating the task of Knowledge-Acquisition (learning from newly generated samples) and Knowledge-Retention (retaining knowledge on previously met samples) as meta-train and meta-test, respectively. Hence, we dub our method as Learning to Retain while Acquiring. Moreover, we identify an implicit aligning factor between the Knowledge-Retention and Knowledge-Acquisition tasks indicating that the proposed student update strategy enforces a common gradient direction for both tasks, alleviating interference between the two objectives. Finally, we support our hypothesis by exhibiting extensive evaluation and comparison of our method with prior arts on multiple datasets.

We introduce a method for novel view synthesis given only a single wide-baseline stereo image pair. In this challenging regime, 3D scene points are regularly observed only once, requiring prior-based reconstruction of scene geometry and appearance. We find that existing approaches to novel view synthesis from sparse observations fail due to recovering incorrect 3D geometry and the high cost of differentiable rendering that precludes their scaling to large-scale training. We take a step towards resolving these shortcomings by formulating a multi-view transformer encoder, proposing an efficient, image-space epipolar line sampling scheme to assemble image features for a target ray, and a lightweight cross-attention-based renderer. Our contributions enable training of our method on a large-scale real-world dataset of indoor and outdoor scenes. In several ablation studies, we demonstrate that our contributions enable learning of powerful multi-view geometry priors while reducing both rendering time and memory footprint. We conduct extensive comparisons on held-out test scenes across two real-world datasets, significantly outperforming prior work on novel view synthesis from sparse image observations and achieving multi-view-consistent novel view synthesis.

We introduce a method that can learn to predict scene-level implicit functions for 3D reconstruction from posed RGBD data. At test time, our system maps a previously unseen RGB image to a 3D reconstruction of a scene via implicit functions. While implicit functions for 3D reconstruction have often been tied to meshes, we show that we can train one using only a set of posed RGBD images. This setting may help 3D reconstruction unlock the sea of accelerometer+RGBD data that is coming with new phones. Our system, D2-DRDF, can match and sometimes outperform current methods that use mesh supervision and shows better robustness to sparse data.

Large scale vision and language models can achieve impressive zero-shot recognition performance by mapping class specific text queries to image content. Two distinct challenges that remain however, are high sensitivity to the choice of handcrafted class names that define queries, and the difficulty of adaptation to new, smaller datasets. Towards addressing these problems, we propose to leverage available data to learn, for each class, an optimal word embedding as a function of the visual content. By learning new word embeddings on an otherwise frozen model, we are able to retain zero-shot capabilities for new classes, easily adapt models to new datasets, and adjust potentially erroneous, non-descriptive or ambiguous class names. We show that our solution can easily be integrated in image classification and object detection pipelines, yields significant performance gains in multiple scenarios and provides insights into model biases and labelling errors.

For point cloud reconstruction-related tasks, the reconstruction losses to evaluate the shape differences between reconstructed results and the ground truths are typically used to train the task networks. Most existing works measure the training loss with point-to-point distance, which may introduce extra defects as predefined matching rules may deviate from the real shape differences. Although some learning-based works have been proposed to overcome the weaknesses of manually-defined rules, they still measure the shape differences in 3D Euclidean space, which may limit their ability to capture defects in reconstructed shapes. In this work, we propose a learning-based Contrastive Adversarial Loss (CALoss) to measure the point cloud reconstruction loss dynamically in a non-linear representation space by combining the contrastive constraint with the adversarial strategy. Specifically, we use the contrastive constraint to help CALoss learn a representation space with shape similarity, while we introduce the adversarial strategy to help CALoss mine differences between reconstructed results and ground truths. According to experiments on reconstruction-related tasks, CALoss can help task networks improve reconstruction performances and learn more representative representations.

We tackle open-world semantic segmentation, which aims at learning to segment arbitrary visual concepts in images, by using only image-text pairs without dense annotations. Existing open-world segmentation methods have shown impressive advances by employing contrastive learning (CL) to learn diverse visual concepts and transferring the learned image-level understanding to the segmentation task. However, these CL-based methods suffer from a train-test discrepancy, since it only considers image-text alignment during training, whereas segmentation requires region-text alignment during testing. In this paper, we proposed a novel Text-grounded Contrastive Learning (TCL) framework that enables a model to directly learn region-text alignment. Our method generates a segmentation mask for a given text, extracts text-grounded image embedding from the masked region, and aligns it with text embedding via TCL. By learning region-text alignment directly, our framework encourages a model to directly improve the quality of generated segmentation masks. In addition, for a rigorous and fair comparison, we present a unified evaluation protocol with widely used 8 semantic segmentation datasets. TCL achieves state-of-the-art zero-shot segmentation performances with large margins in all datasets. Code is available at https://github.com/kakaobrain/tcl.

Scene graph generation (SGG) aims to abstract an image into a graph structure, by representing objects as graph nodes and their relations as labeled edges. However, two knotty obstacles limit the practicability of current SGG methods in real-world scenarios: 1) training SGG models requires time-consuming ground-truth annotations, and 2) the closed-set object categories make the SGG models limited in their ability to recognize novel objects outside of training corpora. To address these issues, we novelly exploit a powerful pre-trained visual-semantic space (VSS) to trigger language-supervised and open-vocabulary SGG in a simple yet effective manner. Specifically, cheap scene graph supervision data can be easily obtained by parsing image language descriptions into semantic graphs. Next, the noun phrases on such semantic graphs are directly grounded over image regions through region-word alignment in the pre-trained VSS. In this way, we enable open-vocabulary object detection by performing object category name grounding with a text prompt in this VSS. On the basis of visually-grounded objects, the relation representations are naturally built for relation recognition, pursuing open-vocabulary SGG. We validate our proposed approach with extensive experiments on the Visual Genome benchmark across various SGG scenarios (i.e., supervised / language-supervised, closed-set / open-vocabulary). Consistent superior performances are achieved compared with existing methods, demonstrating the potential of exploiting pre-trained VSS for SGG in more practical scenarios.

Small on-device models have been successfully trained with user-level differential privacy (DP) for next word prediction and image classification tasks in the past. However, existing methods can fail when directly applied to learn embedding models using supervised training data with a large class space. To achieve user-level DP for large image-to-embedding feature extractors, we propose DP-FedEmb, a variant of federated learning algorithms with per-user sensitivity control and noise addition, to train from user-partitioned data centralized in datacenter. DP-FedEmb combines virtual clients, partial aggregation, private local fine-tuning, and public pretraining to achieve strong privacy utility trade-offs. We apply DP-FedEmb to train image embedding models for faces, landmarks and natural species, and demonstrate its superior utility under same privacy budget on benchmark datasets DigiFace, GLD and iNaturalist. We further illustrate it is possible to achieve strong user-level DP guarantees of epsilon < 2 while controlling the utility drop within 5%, when millions of users can participate in training.

Multi-frame depth estimation generally achieves high accuracy relying on the multi-view geometric consistency. When applied in dynamic scenes, e.g., autonomous driving, this consistency is usually violated in the dynamic areas, leading to corrupted estimations. Many multi-frame methods handle dynamic areas by identifying them with explicit masks and compensating the multi-view cues with monocular cues represented as local monocular depth or features. The improvements are limited due to the uncontrolled quality of the masks and the underutilized benefits of the fusion of the two types of cues. In this paper, we propose a novel method to learn to fuse the multi-view and monocular cues encoded as volumes without needing the heuristically crafted masks. As unveiled in our analyses, the multi-view cues capture more accurate geometric information in static areas, and the monocular cues capture more useful contexts in dynamic areas. To let the geometric perception learned from multi-view cues in static areas propagate to the monocular representation in dynamic areas and let monocular cues enhance the representation of multi-view cost volume, we propose a cross-cue fusion (CCF) module, which includes the cross-cue attention (CCA) to encode the spatially non-local relative intra-relations from each source to enhance the representation of the other. Experiments on real-world datasets prove the significant effectiveness and generalization ability of the proposed method.

The hardware image signal processing (ISP) pipeline is the intermediate layer between the imaging sensor and the downstream application, processing the sensor signal into an RGB image. The ISP is less programmable and consists of a series of processing modules. Each processing module handles a subtask and contains a set of tunable hyperparameters. A large number of hyperparameters form a complex mapping with the ISP output. The industry typically relies on manual and time-consuming hyperparameter tuning by image experts, biased towards human perception. Recently, several automatic ISP hyperparameter optimization methods using downstream evaluation metrics come into sight. However, existing methods for ISP tuning treat the high-dimensional parameter space as a global space for optimization and prediction all at once without inducing the structure knowledge of ISP. To this end, we propose a sequential ISP hyperparameter prediction framework that utilizes the sequential relationship within ISP modules and the similarity among parameters to guide the model sequence process. We validate the proposed method on object detection, image segmentation, and image quality tasks.

Self-supervised learning in vision--language processing (VLP) exploits semantic alignment between imaging and text modalities. Prior work in biomedical VLP has mostly relied on the alignment of single image and report pairs even though clinical notes commonly refer to prior images. This does not only introduce poor alignment between the modalities but also a missed opportunity to exploit rich self-supervision through existing temporal content in the data. In this work, we explicitly account for prior images and reports when available during both training and fine-tuning. Our approach, named BioViL-T, uses a CNN--Transformer hybrid multi-image encoder trained jointly with a text model. It is designed to be versatile to arising challenges such as pose variations and missing input images across time. The resulting model excels on downstream tasks both in single- and multi-image setups, achieving state-of-the-art (SOTA) performance on (I) progression classification, (II) phrase grounding, and (III) report generation, whilst offering consistent improvements on disease classification and sentence-similarity tasks. We release a novel multi-modal temporal benchmark dataset, CXR-T, to quantify the quality of vision--language representations in terms of temporal semantics. Our experimental results show the significant advantages of incorporating prior images and reports to make most use of the data.

Given a piece of text, a video clip and a reference audio, the movie dubbing (also known as visual voice clone, V2C) task aims to generate speeches that match the speaker's emotion presented in the video using the desired speaker voice as reference. V2C is more challenging than conventional text-to-speech tasks as it additionally requires the generated speech to exactly match the varying emotions and speaking speed presented in the video. Unlike previous works, we propose a novel movie dubbing architecture to tackle these problems via hierarchical prosody modeling, which bridges the visual information to corresponding speech prosody from three aspects: lip, face, and scene. Specifically, we align lip movement to the speech duration, and convey facial expression to speech energy and pitch via attention mechanism based on valence and arousal representations inspired by the psychology findings. Moreover, we design an emotion booster to capture the atmosphere from global video scenes. All these embeddings are used together to generate mel-spectrogram, which is then converted into speech waves by an existing vocoder. Extensive experimental results on the V2C and Chem benchmark datasets demonstrate the favourable performance of the proposed method. The code and trained models will be made available at https://github.com/GalaxyCong/HPMDubbing.

Open vocabulary object detection has been greately advanced by the recent development of vision-language pre-trained model, which helps recognizing the novel objects with only semantic categories. The prior works mainly focus on knowledge transferring to the object proposal classification and employ class-agnostic box and mask prediction. In this work, we propose CondHead, a principled dynamic network design to better generalize the box regression and mask segmentation for open vocabulary setting. The core idea is to conditionally parametrize the network heads on semantic embedding and thus the model is guided with class-specific knowledge to better detect novel categories. Specifically, CondHead is composed of two streams of network heads, the dynamically aggregated heads and dynamically generated heads. The former is instantiated with a set of static heads that are conditionally aggregated, these heads are optimized as experts and are expected to learn sophisticated prediction. The Latter is instantiated with dynamically generated parameters and encodes general class-specific information. With such conditional design, the detection model is bridged by the semantic embedding to offer strongly generalizable class-wise box and mask prediction. Our method brings significant improvement to the prior state-of-the-art open vocabulary object detection methods with very minor overhead, e.g., it surpasses a RegionClip model by 3.0 detection AP on novel categories, with only 1.1% more computation.

Detecting mirrors from static images has received significant research interest recently. However, detecting mirrors over dynamic scenes is still under-explored due to the lack of a high-quality dataset and an effective method for video mirror detection (VMD). To the best of our knowledge, this is the first work to address the VMD problem from a deep-learning-based perspective. Our observation is that there are often correspondences between the contents inside (reflected) and outside (real) of a mirror, but such correspondences may not always appear in every frame, e.g., due to the change of camera pose. This inspires us to propose a video mirror detection method, named VMD-Net, that can tolerate spatially missing correspondences by considering the mirror correspondences at both the intra-frame level as well as inter-frame level via a dual correspondence module that looks over multiple frames spatially and temporally for correlating correspondences. We further propose a first large-scale dataset for VMD (named VMD-D), which contains 14,987 image frames from 269 videos with corresponding manually annotated masks. Experimental results show that the proposed method outperforms SOTA methods from relevant fields. To enable real-time VMD, our method efficiently utilizes the backbone features by removing the redundant multi-level module design and gets rid of post-processing of the output maps commonly used in existing methods, making it very efficient and practical for real-time video-based applications. Code, dataset, and models are available at https://jiaying.link/cvpr2023-vmd/

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.

Event cameras sense the intensity changes asynchronously and produce event streams with high dynamic range and low latency. This has inspired research endeavors utilizing events to guide the challenging video super-resolution (VSR) task. In this paper, we make the first at tempt to address a novel problem of achieving VSR at random scales by taking advantages of the high temporal resolution property of events. This is hampered by the difficulties of representing the spatial-temporal information of events when guiding VSR. To this end, we propose a novel framework that incorporates the spatial-temporal interpolation of events to VSR in a unified framework. Our key idea is to learn implicit neural representations from queried spatial-temporal coordinates and features from both RGB frames and events. Our method contains three parts. Specifically, the Spatial-Temporal Fusion (STF) module first learns the 3D features from events and RGB frames. Then, the Temporal Filter (TF) module unlocks more explicit motion information from the events near the queried timestamp and generates the 2D features. Lastly, the Spatial-Temporal Implicit Representation (STIR) module recovers the SR frame in arbitrary resolutions from the outputs of these two modules. In addition, we collect a real-world dataset with spatially aligned events and RGB frames. Extensive experiments show that our method significantly surpass the prior-arts and achieves VSR with random scales, e.g., 6.5. Code and dataset are available at https://.

Answering questions about complex situations in videos requires not only capturing of the presence of actors, objects, and their relations, but also the evolution of these relationships over time. A situation hyper-graph is a representation that describes situations as scene sub-graphs for video frames and hyper-edges for connected sub-graphs, and has been proposed to capture all such information in a compact structured form. In this work, we propose an architecture for Video Question Answering (VQA) that enables answering questions related to video content by predicting situation hyper-graphs, coined Situation Hyper-Graph based Video Question Answering (SHG-VQA). To this end, we train a situation hyper-graph decoder to implicitly identify graph representations with actions and object/human-object relationships from the input video clip and to use cross-attention between the predicted situation hyper-graphs and the question embedding to predict the correct answer. The proposed method is trained in an end-to-end manner and optimized by a cross-entropy based VQA loss function and a Hungarian matching loss for the situation graph prediction. The effectiveness of the proposed architecture is extensively evaluated on two challenging benchmarks: AGQA and STAR. Our results show that learning the underlying situation hyper-graphs helps the system to significantly improve its performance for novel challenges of video question answering task.

Low-light image enhancement (LLIE) investigates how to improve illumination and produce normal-light images. The majority of existing methods improve low-light images via a global and uniform manner, without taking into account the semantic information of different regions. Without semantic priors, a network may easily deviate from a region's original color. To address this issue, we propose a novel semantic-aware knowledge-guided framework (SKF) that can assist a low-light enhancement model in learning rich and diverse priors encapsulated in a semantic segmentation model. We concentrate on incorporating semantic knowledge from three key aspects: a semantic-aware embedding module that wisely integrates semantic priors in feature representation space, a semantic-guided color histogram loss that preserves color consistency of various instances, and a semantic-guided adversarial loss that produces more natural textures by semantic priors. Our SKF is appealing in acting as a general framework in LLIE task. Extensive experiments show that models equipped with the SKF significantly outperform the baselines on multiple datasets and our SKF generalizes to different models and scenes well. The code is available at Semantic-Aware-Low-Light-Image-Enhancement.

3D human body representation learning has received increasing attention in recent years. However, existing works cannot flexibly, controllably and accurately represent human bodies, limited by coarse semantics and unsatisfactory representation capability, particularly in the absence of supervised data. In this paper, we propose a human body representation with fine-grained semantics and high reconstruction-accuracy in an unsupervised setting. Specifically, we establish a correspondence between latent vectors and geometric measures of body parts by designing a part-aware skeleton-separated decoupling strategy, which facilitates controllable editing of human bodies by modifying the corresponding latent codes. With the help of a bone-guided auto-encoder and an orientation-adaptive weighting strategy, our representation can be trained in an unsupervised manner. With the geometrically meaningful latent space, it can be applied to a wide range of applications, from human body editing to latent code interpolation and shape style transfer. Experimental results on public datasets demonstrate the accurate reconstruction and flexible editing abilities of the proposed method. The code will be available at http://cic.tju.edu.cn/faculty/likun/projects/SemanticHuman.