CVPR 2023 Papers

Pretraining visual models on web-scale image-caption datasets has recently emerged as a powerful alternative to traditional pretraining on image classification data. Image-caption datasets are more "open-domain", containing broader scene types and vocabulary words, and result in models that have strong performance in few- and zero-shot recognition tasks. However large-scale classification datasets can provide fine-grained categories with a balanced label distribution. In this work, we study a pretraining strategy that uses both classification and caption datasets to unite their complementary benefits. First, we show that naively unifying the datasets results in sub-optimal performance in downstream zero-shot recognition tasks, as the model is affected by dataset bias: the coverage of image domains and vocabulary words is different in each dataset. We address this problem with novel Prefix Conditioning, a simple yet effective method that helps disentangle dataset biases from visual concepts. This is done by introducing prefix tokens that inform the language encoder of the input data type (e.g., classification vs caption) at training time. Our approach allows the language encoder to learn from both datasets while also tailoring feature extraction to each dataset. Prefix conditioning is generic and can be easily integrated into existing VL pretraining objectives, such as CLIP or UniCL. In experiments, we show that it improves zero-shot image recognition and robustness to image-level distribution shift.

Due to data privacy issues, accelerating networks with tiny training sets has become a critical need in practice. Previous methods mainly adopt filter-level pruning to accelerate networks with scarce training samples. In this paper, we reveal that dropping blocks is a fundamentally superior approach in this scenario. It enjoys a higher acceleration ratio and results in a better latency-accuracy performance under the few-shot setting. To choose which blocks to drop, we propose a new concept namely recoverability to measure the difficulty of recovering the compressed network. Our recoverability is efficient and effective for choosing which blocks to drop. Finally, we propose an algorithm named PRACTISE to accelerate networks using only tiny sets of training images. PRACTISE outperforms previous methods by a significant margin. For 22% latency reduction, PRACTISE surpasses previous methods by on average 7% on ImageNet-1k. It also enjoys high generalization ability, working well under data-free or out-of-domain data settings, too. Our code is at https://github.com/DoctorKey/Practise.

We introduce Power Bundle Adjustment as an expansion type algorithm for solving large-scale bundle adjustment problems. It is based on the power series expansion of the inverse Schur complement and constitutes a new family of solvers that we call inverse expansion methods. We theoretically justify the use of power series and we prove the convergence of our approach. Using the real-world BAL dataset we show that the proposed solver challenges the state-of-the-art iterative methods and significantly accelerates the solution of the normal equation, even for reaching a very high accuracy. This easy-to-implement solver can also complement a recently presented distributed bundle adjustment framework. We demonstrate that employing the proposed Power Bundle Adjustment as a sub-problem solver significantly improves speed and accuracy of the distributed optimization.

Content-aware visual-textual presentation layout aims at arranging spatial space on the given canvas for pre-defined elements, including text, logo, and underlay, which is a key to automatic template-free creative graphic design. In practical applications, e.g., poster designs, the canvas is originally non-empty, and both inter-element relationships as well as inter-layer relationships should be concerned when generating a proper layout. A few recent works deal with them simultaneously, but they still suffer from poor graphic performance, such as a lack of layout variety or spatial non-alignment. Since content-aware visual-textual presentation layout is a novel task, we first construct a new dataset named PKU PosterLayout, which consists of 9,974 poster-layout pairs and 905 images, i.e., non-empty canvases. It is more challenging and useful for greater layout variety, domain diversity, and content diversity. Then, we propose design sequence formation (DSF) that reorganizes elements in layouts to imitate the design processes of human designers, and a novel CNN-LSTM-based conditional generative adversarial network (GAN) is presented to generate proper layouts. Specifically, the discriminator is design-sequence-aware and will supervise the "design" process of the generator. Experimental results verify the usefulness of the new benchmark and the effectiveness of the proposed approach, which achieves the best performance by generating suitable layouts for diverse canvases. The dataset and the source code are available at https://github.com/PKU-ICST-MIPL/PosterLayout-CVPR2023.

Denoising diffusion (score-based) generative models have recently achieved significant accomplishments in generating realistic and diverse data. These approaches define a forward diffusion process for transforming data into noise and a backward denoising process for sampling data from noise. Unfortunately, the generation process of current denoising diffusion models is notoriously slow due to the lengthy iterative noise estimations, which rely on cumbersome neural networks. It prevents the diffusion models from being widely deployed, especially on edge devices. Previous works accelerate the generation process of diffusion model (DM) via finding shorter yet effective sampling trajectories. However, they overlook the cost of noise estimation with a heavy network in every iteration. In this work, we accelerate generation from the perspective of compressing the noise estimation network. Due to the difficulty of retraining DMs, we exclude mainstream training-aware compression paradigms and introduce post-training quantization (PTQ) into DM acceleration. However, the output distributions of noise estimation networks change with time-step, making previous PTQ methods fail in DMs since they are designed for single-time step scenarios. To devise a DM-specific PTQ method, we explore PTQ on DM in three aspects: quantized operations, calibration dataset, and calibration metric. We summarize and use several observations derived from all-inclusive investigations to formulate our method, which especially targets the unique multi-time-step structure of DMs. Experimentally, our method can directly quantize full-precision DMs into 8-bit models while maintaining or even improving their performance in a training-free manner. Importantly, our method can serve as a plug-and-play module on other fast-sampling methods, such as DDIM.

Existing Temporal Action Detection (TAD) methods typically take a pre-processing step in converting an input varying-length video into a fixed-length snippet representation sequence, before temporal boundary estimation and action classification. This pre-processing step would temporally downsample the video, reducing the inference resolution and hampering the detection performance in the original temporal resolution. In essence, this is due to a temporal quantization error introduced during the resolution downsampling and recovery. This could negatively impact the TAD performance, but is largely ignored by existing methods. To address this problem, in this work we introduce a novel model-agnostic post-processing method without model redesign and retraining. Specifically, we model the start and end points of action instances with a Gaussian distribution for enabling temporal boundary inference at a sub-snippet level. We further introduce an efficient Taylor-expansion based approximation, dubbed as Gaussian Approximated Post-processing (GAP). Extensive experiments demonstrate that our GAP can consistently improve a wide variety of pre-trained off-the-shelf TAD models on the challenging ActivityNet (+0.2% 0.7% in average mAP) and THUMOS (+0.2% 0.5% in average mAP) benchmarks. Such performance gains are already significant and highly comparable to those achieved by novel model designs. Also, GAP can be integrated with model training for further performance gain. Importantly, GAP enables lower temporal resolutions for more efficient inference, facilitating low-resource applications. The code is available in https://github.com/sauradip/GAP.

The CLIP model has been recently proven to be very effective for a variety of cross-modal tasks, including the evaluation of captions generated from vision-and-language architectures. In this paper, we propose a new recipe for a contrastive-based evaluation metric for image captioning, namely Positive-Augmented Contrastive learning Score (PAC-S), that in a novel way unifies the learning of a contrastive visual-semantic space with the addition of generated images and text on curated data. Experiments spanning several datasets demonstrate that our new metric achieves the highest correlation with human judgments on both images and videos, outperforming existing reference-based metrics like CIDEr and SPICE and reference-free metrics like CLIP-Score. Finally, we test the system-level correlation of the proposed metric when considering popular image captioning approaches, and assess the impact of employing different cross-modal features. Our source code and trained models are publicly available at: https://github.com/aimagelab/pacscore.

Vision-Language Pre-Training (VLP) has shown promising capabilities to align image and text pairs, facilitating a broad variety of cross-modal learning tasks. However, we observe that VLP models often lack the visual grounding/localization capability which is critical for many downstream tasks such as visual reasoning. In this work, we propose a novel Position-guided Text Prompt (PTP) paradigm to enhance the visual grounding ability of cross-modal models trained with VLP. Specifically, in the VLP phase, PTP divides the image into NxN blocks, and identifies the objects in each block through the widely used object detector in VLP. It then reformulates the visual grounding task into a fill-in-the-blank problem given a PTP by encouraging the model to predict the objects in the given blocks or regress the blocks of a given object, e.g. filling "P" or "O" in a PTP "The block P has a O". This mechanism improves the visual grounding capability of VLP models and thus helps them better handle various downstream tasks. By introducing PTP into several state-of-the-art VLP frameworks, we observe consistently significant improvements across representative cross-modal learning model architectures and several benchmarks, e.g. zero-shot Flickr30K Retrieval (+4.8 in average recall@1) for ViLT baseline, and COCO Captioning (+5.3 in CIDEr) for SOTA BLIP baseline. Moreover, PTP achieves comparable results with object-detector based methods, and much faster inference speed since PTP discards its object detector for inference while the later cannot. Our code and pre-trained weight will be released.

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.

Human pose and shape (HPS) estimation methods achieve remarkable results. However, current HPS benchmarks are mostly designed to test models in scenarios that are similar to the training data. This can lead to critical situations in real-world applications when the observed data differs significantly from the training data and hence is out-of-distribution (OOD). It is therefore important to test and improve the OOD robustness of HPS methods. To address this fundamental problem, we develop a simulator that can be controlled in a fine-grained manner using interpretable parameters to explore the manifold of images of human pose, e.g. by varying poses, shapes, and clothes. We introduce a learning-based testing method, termed PoseExaminer, that automatically diagnoses HPS algorithms by searching over the parameter space of human pose images to find the failure modes. Our strategy for exploring this high-dimensional parameter space is a multi-agent reinforcement learning system, in which the agents collaborate to explore different parts of the parameter space. We show that our PoseExaminer discovers a variety of limitations in current state-of-the-art models that are relevant in real-world scenarios but are missed by current benchmarks. For example, it finds large regions of realistic human poses that are not predicted correctly, as well as reduced performance for humans with skinny and corpulent body shapes. In addition, we show that fine-tuning HPS methods by exploiting the failure modes found by PoseExaminer improve their robustness and even their performance on standard benchmarks by a significant margin. The code are available for research purposes.

Self-supervised facial representation has recently attracted increasing attention due to its ability to perform face understanding without relying on large-scale annotated datasets heavily. However, analytically, current contrastive-based self-supervised learning (SSL) still performs unsatisfactorily for learning facial representation. More specifically, existing contrastive learning (CL) tends to learn pose-invariant features that cannot depict the pose details of faces, compromising the learning performance. To conquer the above limitation of CL, we propose a novel Pose-disentangled Contrastive Learning (PCL) method for general self-supervised facial representation. Our PCL first devises a pose-disentangled decoder (PDD) with a delicately designed orthogonalizing regulation, which disentangles the pose-related features from the face-aware features; therefore, pose-related and other pose-unrelated facial information could be performed in individual subnetworks and do not affect each other's training. Furthermore, we introduce a pose-related contrastive learning scheme that learns pose-related information based on data augmentation of the same image, which would deliver more effective face-aware representation for various downstream tasks. We conducted linear evaluation on four challenging downstream facial understanding tasks, i.e., facial expression recognition, face recognition, AU detection and head pose estimation.Experimental results demonstrate that PCL significantly outperforms cutting-edge SSL methods. Our Code is available at https://github.com/DreamMr/PCL.

Pose synchronization, which seeks to estimate consistent absolute poses among a collection of objects from noisy relative poses estimated between pairs of objects in isolation, is a fundamental problem in many inverse applications. This paper studies an extreme setting where multiple relative pose estimates exist between each object pair, and the majority is incorrect. Popular methods that solve pose synchronization via recovering a low-rank matrix that encodes relative poses in block fail under this extreme setting. We introduce a three-step algorithm for pose synchronization under multiple relative pose inputs. The first step performs diffusion and clustering to compute the candidate poses of the input objects. We present a theoretical result to justify our diffusion formulation. The second step jointly optimizes the best pose for each object. The final step refines the output of the second step. Experimental results on benchmark datasets of structurefrom-motion and scan-based geometry reconstruction show that our approach offers more accurate absolute poses than state-of-the-art pose synchronization techniques.

Implicit neural representations (INR) have gained significant popularity for signal and image representation for many end-tasks, such as superresolution, 3D modeling, and more. Most INR architectures rely on sinusoidal positional encoding, which accounts for high-frequency information in data. However, the finite encoding size restricts the model's representational power. Higher representational power is needed to go from representing a single given image to representing large and diverse datasets. Our approach addresses this gap by representing an image with a polynomial function and eliminates the need for positional encodings. Therefore, to achieve a progressively higher degree of polynomial representation, we use element-wise multiplications between features and affine-transformed coordinate locations after every ReLU layer. The proposed method is evaluated qualitatively and quantitatively on large datasets like ImageNet. The proposed Poly-INR model performs comparably to state-of-the-art generative models without any convolution, normalization, or self-attention layers, and with far fewer trainable parameters. With much fewer training parameters and higher representative power, our approach paves the way for broader adoption of INR models for generative modeling tasks in complex domains. The code is available at https://github.com/Rajhans0/Poly_INR

In this work, instead of directly predicting the pixel-level segmentation masks, the problem of referring image segmentation is formulated as sequential polygon generation, and the predicted polygons can be later converted into segmentation masks. This is enabled by a new sequence-to-sequence framework, Polygon Transformer (PolyFormer), which takes a sequence of image patches and text query tokens as input, and outputs a sequence of polygon vertices autoregressively. For more accurate geometric localization, we propose a regression-based decoder, which predicts the precise floating-point coordinates directly, without any coordinate quantization error. In the experiments, PolyFormer outperforms the prior art by a clear margin, e.g., 5.40% and 4.52% absolute improvements on the challenging RefCOCO+ and RefCOCOg datasets. It also shows strong generalization ability when evaluated on the referring video segmentation task without fine-tuning, e.g., achieving competitive 61.5% J&F on the Ref-DAVIS17 dataset.

In this work, we show that it is feasible to perform multiple tasks concurrently on point cloud with a straightforward yet effective multi-task network. Our framework, Poly-PC, tackles the inherent obstacles (e.g., different model architectures caused by task bias and conflicting gradients caused by multiple dataset domains, etc.) of multi-task learning on point cloud. Specifically, we propose a residual set abstraction (Res-SA) layer for efficient and effective scaling in both width and depth of the network, hence accommodating the needs of various tasks. We develop a weight-entanglement-based one-shot NAS technique to find optimal architectures for all tasks. Moreover, such technique entangles the weights of multiple tasks in each layer to offer task-shared parameters for efficient storage deployment while providing ancillary task-specific parameters for learning task-related features. Finally, to facilitate the training of Poly-PC, we introduce a task-prioritization-based gradient balance algorithm that leverages task prioritization to reconcile conflicting gradients, ensuring high performance for all tasks. Benefiting from the suggested techniques, models optimized by Poly-PC collectively for all tasks keep fewer total FLOPs and parameters and outperform previous methods. We also demonstrate that Poly-PC allows incremental learning and evades catastrophic forgetting when tuned to a new task.

Recent progress on vision-language foundation models have brought significant advancement to building general-purpose robots. By using the pre-trained models to encode the scene and instructions as inputs for decision making, the instruction-conditioned policy can generalize across different objects and tasks. While this is encouraging, the policy still fails in most cases given an unseen task or environment. In this work, we propose Policy Adaptation from Foundation model Feedback (PAFF). When deploying the trained policy to a new task or a new environment, we first let the policy play with randomly generated instructions to record the demonstrations. While the execution could be wrong, we can use the pre-trained foundation models to provide feedback to relabel the demonstrations. This automatically provides new pairs of demonstration-instruction data for policy fine-tuning. We evaluate our method on a broad range of experiments with the focus on generalization on unseen objects, unseen tasks, unseen environments, and sim-to-real transfer. We show PAFF improves baselines by a large margin in all cases.

Single-chip polarized color photography provides both visual textures and object surface information in one snapshot. However, the use of an additional directional polarizing filter array tends to lower photon count and SNR, when compared to conventional color imaging. As a result, such a bilayer structure usually leads to unpleasant noisy images and undermines performance of polarization analysis, especially in low-light conditions. It is a challenge for traditional image processing pipelines owing to the fact that the physical constraints exerted implicitly in the channels are excessively complicated. In this paper, we propose to tackle this issue through a noise modeling method for realistic data synthesis and a powerful network structure inspired by vision Transformer. A real-world polarized color image dataset of paired raw short-exposed noisy images and long-exposed reference images is captured for experimental evaluation, which has demonstrated the effectiveness of our approaches for data synthesis and polarized color image denoising.

Indirect time-of-flight (iToF) imaging allows us to capture dense depth information at a low cost. However, iToF imaging often suffers from multipath interference (MPI) artifacts in the presence of scattering media, resulting in severe depth-accuracy degradation. For instance, iToF cameras cannot measure depth accurately through fog because ToF active illumination scatters back to the sensor before reaching the farther target surface. In this work, we propose a polarimetric iToF imaging method that can capture depth information robustly through scattering media. Our observations on the principle of indirect ToF imaging and polarization of light allow us to formulate a novel computational model of scattering-aware polarimetric phase measurements that enables us to correct MPI errors. We first devise a scattering-aware polarimetric iToF model that can estimate the phase of unpolarized backscattered light. We then combine the optical filtering of polarization and our computational modeling of unpolarized backscattered light via scattering analysis of phase and amplitude. This allows us to tackle the MPI problem by estimating the scattering energy through the participating media. We validate our method on an experimental setup using a customized off-the-shelf iToF camera. Our method outperforms baseline methods by a significant margin by means of our scattering model and polarimetric phase measurements.

We propose a novel method that renders point clouds as if they are surfaces. The proposed method is differentiable and requires no scene-specific optimization. This unique capability enables, out-of-the-box, surface normal estimation, rendering room-scale point clouds, inverse rendering, and ray tracing with global illumination. Unlike existing work that focuses on converting point clouds to other representations--e.g., surfaces or implicit functions--our key idea is to directly infer the intersection of a light ray with the underlying surface represented by the given point cloud. Specifically, we train a set transformer that, given a small number of local neighbor points along a light ray, provides the intersection point, the surface normal, and the material blending weights, which are used to render the outcome of this light ray. Localizing the problem into small neighborhoods enables us to train a model with only 48 meshes and apply it to unseen point clouds. Our model achieves higher estimation accuracy than state-of-the-art surface reconstruction and point-cloud rendering methods on three test sets. When applied to room-scale point clouds, without any scene-specific optimization, the model achieves competitive quality with the state-of-the-art novel-view rendering methods. Moreover, we demonstrate ability to render and manipulate Lidar-scanned point clouds such as lighting control and object insertion.