ECCV 2022 论文列表

"In this paper, we mainly focus on the problem of how to learn additional feature representations for few-shot image classification through pretext tasks (e.g., rotation or color permutation and so on). This additional knowledge generated by pretext tasks can further improve the performance of few-shot learning (FSL) as it differs from human-annotated supervision (i.e., class labels of FSL tasks). To solve this problem, we present a plug-in Hierarchical Tree Structure-aware (HTS) method, which not only learns the relationship of FSL and pretext tasks, but more importantly, can adaptively select and aggregate feature representations generated by pretext tasks to maximize the performance of FSL tasks. A hierarchical tree constructing component and a gated selection aggregating component is introduced to construct the tree structure and find richer transferable knowledge that can rapidly adapt to novel classes with a few labeled images. Extensive experiments show that our HTS can significantly enhance multiple few-shot methods to achieve new state-of-the-art performance on four benchmark datasets. The code is available at: https://github.com/remiMZ/HTS-ECCV22."

"Do deep learning models for instance segmentation generalize to novel objects in a systematic way? For classification, such behavior has been questioned. In this study, we aim to understand if certain design decisions such as framework, architecture or pre-training contribute to the semantic understanding of instance segmentation. To answer this question, we consider a special case of robustness and compare pre-trained models on a challenging benchmark for object-centric out-of-distribution texture. We do not introduce another method in this work. Instead, we take a step back and evaluate a broad range of existing literature. This includes Cascade and Mask R-CNN, Swin Transformer, BMask, YOLACT(++), DETR, BCNet, SOTR and SOLOv2. We find that YOLACT++, SOTR and SOLOv2 are significantly more robust to out-of-distribution texture than other frameworks. In addition, we show that deeper and dynamic architectures improve robustness whereas training schedules, data augmentation and pre-training have only a minor impact. In summary we evaluate 68 models on 61 versions of MS COCO for a total of 4148 evaluations."

"We study the problem of translating an image-based, step-by-step assembly manual created by human designers into machine-interpretable instructions. We formulate this problem as a sequential prediction task: at each step, our model reads the manual, locates the components to be added to the current shape, and infers their 3D poses. This task poses the challenge of establishing a 2D-3D correspondence between the manual image and the real 3D object, and 3D pose estimation for unseen 3D objects, since a new component to be added in a step can be an object built from previous steps. To address these two challenges, we present a novel learning-based framework, the Manual-to-Executable-Plan Network (MEPNet), which reconstructs the assembly steps from a sequence of manual images. The key idea is to integrate neural 2D keypoint detection modules and 2D-3D projection algorithms for high-precision prediction and strong generalization to unseen components. The MEPNet outperforms existing methods on three newly collected LEGO manual datasets and a Minecraft house dataset."

"Implicit Neural Representations (INRs) have emerged and shown their benefits over discrete representations in recent years. However, fitting an INR to the given observations usually requires optimization with gradient descent from scratch, which is inefficient and does not generalize well with sparse observations. To address this problem, most of the prior works train a hypernetwork that generates a single vector to modulate the INR weights, where the single vector becomes an information bottleneck that limits the reconstruction precision of the output INR. Recent work shows that the whole set of weights in INR can be precisely inferred without the single-vector bottleneck by gradient-based meta-learning. Motivated by a generalized formulation of gradient-based meta-learning, we propose a formulation that uses Transformers as hypernetworks for INRs, where it can directly build the whole set of INR weights with Transformers specialized as set-to-set mapping. We demonstrate the effectiveness of our method for building INRs in different tasks and domains, including 2D image regression and view synthesis for 3D objects. Our work draws connections between the Transformer hypernetworks and gradient-based meta-learning algorithms and we provide further analysis for understanding the generated INRs."

"Particle-based systems provide a flexible and unified way to simulate physics systems with complex dynamics. Most existing data-driven simulators for particle-based systems adopt graph neural networks (GNNs) as their network backbones, as particles and their interactions can be naturally represented by graph nodes and graph edges. However, while particle-based systems usually contain hundreds even thousands of particles, the explicit modeling of particle interactions as graph edges inevitably leads to a significant computational overhead, due to the increased number of particle interactions. Consequently, in this paper we propose a novel Transformer-based method, dubbed as Transformer with Implicit Edges (TIE), to capture the rich semantics of particle interactions in an edge-free manner. The core idea of TIE is to decentralize the computation involving pair-wise particle interactions into per-particle updates. This is achieved by adjusting the self-attention module to resemble the update formula of graph edges in GNN. To improve the generalization ability of TIE, we further amend TIE with learnable material-specific abstract particles to disentangle global material-wise semantics from local particle-wise semantics. We evaluate our model on diverse domains of varying complexity and materials. Compared with existing GNN-based methods, without bells and whistles, TIE achieves superior performance and generalization across all these domains. Codes and models are available at https://github.com/ftbabi/TIE_ECCV2022.git."

"We propose a trainable Image Signal Processing (ISP) framework that produces DSLR quality images given RAW images captured by a smartphone. To address the color misalignments between training image pairs, we employ a color-conditional ISP network and optimize a novel parametric color mapping between each input RAW and reference DSLR image. During inference, we predict the target color image by designing a color prediction network with efficient Global Context Transformer modules. The latter effectively leverage global information to learn consistent color and tone mappings. We further propose a robust masked aligned loss to identify and discard regions with inaccurate motion estimation during training. Lastly, we introduce the ISP in the Wild (ISPW) dataset, consisting of weakly paired phone RAW and DSLR sRGB images. We extensively evaluate our method, setting a new state-of-the-art on two datasets. The code is available at https://github.com/4rdhendu/TransformPhone2DSLR."

"This work investigates the entanglement between Continual Learning (CL) and Transfer Learning (TL). In particular, we shed light on the widespread application of network pretraining, highlighting that it is itself subject to catastrophic forgetting. Unfortunately, this issue leads to the under-exploitation of knowledge transfer during later tasks. On this ground, we propose Transfer without Forgetting (TwF), a hybrid approach building upon a fixed pretrained sibling network, which continuously propagates the knowledge inherent in the source domain through a layer-wise loss term. Our experiments indicate that TwF steadily outperforms other CL methods across a variety of settings, averaging a 4.81% gain in Class-Incremental accuracy over a variety of datasets and different buffer sizes."

"This paper presents a transformer framework for few-shot learning, termed TransVLAD, with one focus showing the power of locally aggregated descriptors for few-shot learning. Our TransVLAD model is simple: a standard transformer encoder following a NeXtVLAD aggregation module to output the locally aggregated descriptors. In contrast to the prevailing use of CNN as part of the feature extractor, we are the first to prove self-supervised learning like masked autoencoders (MAE) can deal with the overfitting of transformers in few-shot image classification. Besides, few-shot learning can benefit from this general-purpose pre-training. Then, we propose two methods to mitigate few-shot biases, supervision bias and simple-characteristic bias. The first method is introducing masking operation into fine-tuning, by which we accelerate fine-tuning (by more than 3x) and improve accuracy. The second one is adapting focal loss into soft focal loss to focus on hard characteristics learning. Our TransVLAD finally tops 10 benchmarks on five popular few-shot datasets by an average of more than 2%."

"Image matting refers to predicting the alpha values of unknown foreground areas from natural images. Prior methods have focused on propagating alpha values from known to unknown regions. However, not all natural images have a specifically known foreground. Images of transparent objects, like glass, smoke, web, etc., have less or no known foreground. In this paper, we propose a Transformer-based network, TransMatting, to model transparent objects with a big receptive field. Specifically, we redesign the trimap as three learnable tri-tokens for introducing advanced semantic features into the self-attention mechanism. A small convolutional network is proposed to utilize the global feature and non-background mask to guide the multi-scale feature propagation from encoder to decoder for maintaining the contexture of transparent objects. In addition, we create a high-resolution matting dataset of transparent objects with small known foreground areas. Experiments on several matting benchmarks demonstrate the superiority of our proposed method over the current state-of-the-art methods."

"Grasp pose estimation is an important issue for robots to interact with the real world. However, most of existing methods require exact 3D object models available beforehand or a large amount of grasp annotations for training. To avoid these problems, we propose TransGrasp, a category-level grasp pose estimation method that predicts grasp poses of a category of objects by labeling only one object instance. Specifically, we perform grasp pose transfer across a category of objects based on their shape correspondences and propose a grasp pose refinement module to further fine-tune grasp pose of grippers so as to ensure successful grasps. Experiments demonstrate the effectiveness of our method on achieving high-quality grasps with the transferred grasp poses. Our code is available at https://github.com/yanjh97/TransGrasp."

"Unsupervised semantic segmentation aims to obtain high-level semantic representation on low-level visual features without manual annotations. Most existing methods are bottom-up approaches that try to group pixels into regions based on their visual cues or certain predefined rules. As a result, it is difficult for these bottom-up approaches to generate fine-grained semantic segmentation when coming to complicated scenes with multiple objects and some objects sharing similar visual appearance. In contrast, we propose the first top-down unsupervised semantic segmentation framework for fine-grained segmentation in extremely complicated scenarios. Specifically, we first obtain rich high-level structured semantic concept information from large-scale vision data in a self-supervised learning manner, and use such information as a prior to discover potential semantic categories presented in target datasets. Secondly, the discovered high-level semantic categories are mapped to low-level pixel features by calculating the class activate map (CAM) with respect to certain discovered semantic representation. Lastly, the obtained CAMs serve as pseudo labels to train the segmentation module and produce final semantic segmentation. Experimental results on multiple semantic segmentation benchmarks show that our top-down unsupervised segmentation is robust to both object-centric and scene-centric datasets under different settings of semantic granularity level, and outperforms all the current state-of-the-art bottom-up methods."

"Vision Transformers (ViTs) is emerging as an alternative to convolutional neural networks (CNNs) for visual recognition. They achieve competitive results with CNNs but the lack of the typical convolutional inductive bias makes them more data-hungry than common CNNs. They are often pretrained on JFT-300M or at least ImageNet and few works study training ViTs with limited data. In this paper, we investigate how to train ViTs with limited data (e.g., 2040 images). We give theoretical analyses that our method (based on parametric instance discrimination) is superior to other methods in that it can capture both feature alignment and instance similarities. We achieve state-of-the-art results when training from scratch on 7 small datasets under various ViT backbones. We also investigate the transferring ability of small datasets and find that representations learned from small datasets can even improve large-scale ImageNet training."

"It is well noted that coordinate-based MLPs benefit---in terms of preserving high-frequency information---through the encoding of coordinate positions as an array of Fourier features. Hitherto, the rationale for the effectiveness of these \emph{positional encodings} has been mainly studied through a Fourier lens. In this paper, we strive to broaden this understanding by showing that alternative non-Fourier embedding functions can indeed be used for positional encoding. Moreover, we show that their performance is entirely determined by a trade-off between the stable rank of the embedded matrix and the distance preservation between embedded coordinates. We further establish that the now ubiquitous Fourier feature mapping of position is a special case that fulfills these conditions. Consequently, we present a more general theory to analyze positional encoding in terms of shifted basis functions. In addition, we argue that employing a more complex positional encoding---that scales exponentially with the number of modes---requires only a linear (rather than deep) coordinate function to achieve comparable performance. Counter-intuitively, we demonstrate that trading positional embedding complexity for network deepness is orders of magnitude faster than current state-of-the-art; despite the additional embedding complexity. To this end, we develop the necessary theoretical formulae and empirically verify that our theoretical claims hold in practice."

"The tracking-by-detection paradigm today has become the dominant method for multi-object tracking and works by detecting objects in each frame and then performing data association across frames. However, its sequential frame-wise matching property fundamentally suffers from the intermediate interruptions in a video, such as object occlusions, fast camera movements, and abrupt light changes. Moreover, it typically overlooks temporal information beyond the two frames for matching. In this paper, we investigate an alternative by treating object association as clip-wise matching. Our new perspective views a single long video sequence as multiple short clips, and then the tracking is performed both within and between the clips. The benefits of this new approach are two folds. First, our method is robust to tracking error accumulation or propagation, as the video chunking allows bypassing the interrupted frames, and the short clip tracking avoids the conventional error-prone long-term track memory management. Second, the multiple frame information is aggregated during the clip-wise matching, resulting in a more accurate long-range track association than the current frame-wise matching. Given the state-of-the-art tracking-by-detection tracker, QDTrack, we showcase how the tracking performance improves with our new tracking formulation. In addition, we designed a novel Transformer-based clip tracker tailored for the effective inter-clip association. We evaluate our proposals on two tracking benchmarks, TAO and MOT17 that have complementary characteristics and challenges each other. Our new tracking formulation greatly improves association quality, leading to a new state-of-the-art in TAO."

"Multi-object tracking (MOT) requires detecting and associating objects through frames. Unlike tracking via detected bounding boxes or center points, we propose tracking objects as pixel-wise distributions. We instantiate this idea on a transformer-based architecture named P3AFormer, with pixel-wise propagation, prediction, and association. P3AFormer propagates pixel-wise features guided by flow information to pass messages between frames. Further, P3AFormer adopts a meta-architecture to produce multi-scale object feature maps. During inference, a pixel-wise association procedure is proposed to recover object connections through frames based on the pixel-wise prediction. P3AFormer yields 81.2\% in terms of MOTA on the MOT17 benchmark -- highest among all transformer networks to reach 80\% MOTA in literature. P3AFormer also outperforms state-of-the-arts on the MOT20 and KITTI benchmarks. The code is at https://github.com/dvlab-research/ECCV22-P3AFormer-Tracking-Objects-as-Pixel-wise-Distributions."

"Current multi-category Multiple Object Tracking (MOT) metrics use class labels to group tracking results for per-class evaluation. Similarly, MOT methods typically only associate objects with the same class predictions. These two prevalent strategies in MOT implicitly assume that the classification performance is near-perfect. However, this is far from the case in recent large-scale MOT datasets, which contain large numbers of classes with many rare or semantically similar categories. Therefore, the resulting inaccurate classification leads to sub-optimal tracking and inadequate benchmarking of trackers. We address these issues by disentangling classification from tracking. We introduce a new metric, Track Every Thing Accuracy (TETA), breaking tracking measurement into three sub-factors: localization, association, and classification, allowing comprehensive benchmarking of tracking performance even under inaccurate classification. TETA also deals with the challenging incomplete annotation problem in large-scale tracking datasets. We further introduce a Track Every Thing tracker (TETer), that performs association using Class Exemplar Matching (CEM). Our experiments show that TETA evaluates trackers more comprehensively, and TETer achieves significant improvements on the challenging large-scale datasets BDD100K and TAO compared to the state-of-the-art."

"Text to Image generation is a fundamental and inevitable challenging task for visual linguistic modeling. The recent surge in this area such as DALL·E has shown breathtaking technical breakthroughs, however, it still lacks a precise control of the spatial relation corresponding to semantic text. To tackle this problem, mouse trace paired with text provides an interactive way, in which users can describe the imagined image with natural language while drawing traces to locate those they want. However, this brings the challenges of both controllability and compositionality of the generation. Motivated by this, we propose a Trace Controlled Text to Image Generation model (TCTIG), which takes trace as a bridge between semantic concepts and spatial conditions. Moreover, we propose a set of new technique to enhance the controllability and compositionality of generation, including trace guided re-weighting loss (TGR) and semantic aligned augmentation (SAA). In addition, we establish a solid benchmark for the trace-controlled text-to-image generation task, and introduce several new metrics to evaluate both the controllability and compositionality of the model. Upon that, we demonstrate TCTIG’s superior performance and further present the fruitful qualitative analysis of our model."

"Facial pose estimation refers to the task of predicting face orientation from a single RGB image. It is an important research topic with a wide range of applications in computer vision. Label distribution learning (LDL) based methods have been recently proposed for facial pose estimation, which achieve promising results. However, there are two major issues in existing LDL methods. First, the expectations of label distributions are biased, leading to a biased pose estimation. Second, fixed distribution parameters are applied for all learning samples, severely limiting the model capability. In this paper, we propose an Anisotropic Spherical Gaussian (ASG)-based LDL approach for facial pose estimation. In particular, our approach adopts the spherical Gaussian distribution on a unit sphere which constantly generates unbiased expectation. Meanwhile, we introduce a new loss function that allows the network to learn the distribution parameter for each learning sample flexibly. Extensive experimental results show that our method sets new state-of-the-art records on AFLW2000 and BIWI datasets."

"Spiking Neural Networks (SNNs) can be energy efficient alternatives to commonly used deep neural networks (DNNs). However, computation over multiple timesteps increases latency and energy and incurs memory access overhead of membrane potentials. Hence, latency reduction is pivotal to obtain SNNs with high energy efficiency. But, reducing latency can have an adverse effect on accuracy. To optimize the accuracy-energy-latency trade-off, we propose a temporal pruning method which starts with an SNN of T timesteps, and reduces T every iteration of training, with threshold and leak as trainable parameters. This results in a continuum of SNNs from T timesteps, all the way up to unit timestep. Training SNNs directly with 1 timestep results in convergence failure due to layerwise spike vanishing and difficulty in finding optimum thresholds. The proposed temporal pruning overcomes this by enabling the learning of suitable layerwise thresholds with backpropagation by maintaining sufficient spiking activity. Using the proposed algorithm, we achieve top-1 accuracy of 93.05%, 70.15% and 69.00% on CIFAR-10, CIFAR-100 and ImageNet, respectively with VGG16, in just 1 timestep. Note, SNNs with leaky-integrate-and-fire (LIF) neurons behave as Recurrent Neural Networks (RNNs), with the membrane potential retaining information of previous inputs. The proposed SNNs also enable performing sequential tasks such as reinforcement learning on Cartpole and Atari pong environments using only 1 to 5 timesteps."