Papers

There are two key issues that limit further improvements in the performance of existing rotational detectors: 1)Periodic sudden change of the parameters in the rotating bounding box (RBBox) definition causes a numerical discontinuity in the loss (such as smoothL1 loss). 2)There is a gap of optimization asynchrony between the loss in the RBBox regression and evaluation metrics. In this paper, we define a new distance formulation between two convex polygons describing the overlapping degree and non-overlapping degree. Based on this smooth distance, we propose a loss called Polygon-to-Polygon distance loss (P2P Loss). The distance is derived from the area sum of triangles specified by the vertexes of one polygon and the edges of the other. Therefore, the P2P Loss is continuous, differentiable, and inherently free from any RBBox definition. Our P2P Loss is not only consistent with the detection metrics but also able to measure how far, as well as how similar, an RBBox is from another one even when they are completely non-overlapping. These features allow the RetinaNet using the P2P Loss to achieve 79.15% mAP on the DOTA dataset, which is quite competitive compared with many state-of-the-art rotated object detectors.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Video object detection has been an important yet challenging topic in computer vision. Traditional methods mainly focus on designing the image-level or box-level feature propagation strategies to exploit temporal information. This paper argues that with a more effective and efficient feature propagation framework, video object detectors can gain improvement in terms of both accuracy and speed. For this purpose, this paper studies object-level feature propagation, and proposes an object query propagation (QueryProp) framework for high-performance video object detection. The proposed QueryProp contains two propagation strategies: 1) query propagation is performed from sparse key frames to dense non-key frames to reduce the redundant computation on non-key frames; 2) query propagation is performed from previous key frames to the current key frame to improve feature representation by temporal context modeling. To further facilitate query propagation, an adaptive propagation gate is designed to achieve flexible key frame selection. We conduct extensive experiments on the ImageNet VID dataset. QueryProp achieves comparable accuracy with state-of-the-art methods and strikes a decent accuracy/speed trade-off.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Recent self-supervised video representation learning methods have found significant success by exploring essential properties of videos, e.g. speed, temporal order, etc. This work exploits an essential yet under-explored property of videos, the \textit{video continuity}, to obtain supervision signals in self-supervised representation learning. Specifically, we formulate three novel continuity-related pretext tasks, i.e. continuity justification, discontinuity localization, and missing section approximation, that supervise a shared backbone for video representation learning. This self-supervision approach, termed as Continuity Perception Network (CPNet), encourages its backbone network to learn local and long-ranged motion and context representations and outperforms prior arts on multiple downstream tasks, such as action recognition, video retrieval, and action localization. Also, the video continuity can be complementary to other video properties for representation learning, and integrating the proposed pretext task to prior arts can yield much performance gains.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Due to high-speed motion blur and challenging illumination, conventional frame-based cameras have encountered an important challenge in object detection tasks. Neuromorphic cameras that output asynchronous visual streams instead of intensity frames, by taking the advantage of high temporal resolution and high dynamic range, have brought a new perspective to address the challenge. In this paper, we propose a novel problem setting, retinomorphic object detection, which is the first trial that integrates foveal-like and peripheral-like visual streams. Technically, we first build a large-scale multimodal neuromorphic object detection dataset (i.e., PKU-Vidar-DVS) over 215.5k spatio-temporal synchronized labels. Then, we design temporal aggregation representations to preserve the spatio-temporal information from asynchronous visual streams. Finally, we present a novel bio-inspired unifying framework to fuse two sensing modalities via a dynamic interaction mechanism. Our experimental evaluation shows that our approach has significant improvements over the state-of-the-art methods with the single-modality, especially in high-speed motion and low-light scenarios. We hope that our work will attract further research into this newly identified, yet crucial research direction. Our dataset can be available at https://www.pkuml.org/resources/pku-vidar-dvs.html.

poster session 1 @ red 3, poster session 11 @ red 3, oral session 11 @ red 3, poster session 1, poster session 11, oral session 11

Large capacity deep learning models are often prone to a high generalization gap when trained with a limited amount of labeled training data. A recent class of methods to address this problem uses various ways to construct a new training sample by mixing a pair (or more) of training samples. We propose PatchUp, a hidden state block-level regularization technique for Convolutional Neural Networks (CNNs), that is applied on selected contiguous blocks of feature maps from a random pair of samples. Our approach improves the robustness of CNN models against the manifold intrusion problem that may occur in other state-of-the-art mixing approaches. Moreover, since we are mixing the contiguous block of features in the hidden space, which has more dimensions than the input space, we obtain more diverse samples for training towards different dimensions. Our experiments on CIFAR10/100, SVHN, Tiny-ImageNet, and ImageNet using ResNet architectures including PreActResnet18/34, WRN-28-10, ResNet101/152 models show that PatchUp improves upon, or equals, the performance of current state-of-the-art regularizers for CNNs. We also show that PatchUp can provide a better generalization to deformed samples and is more robust against adversarial attacks.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Hierarchical structures are popular in recent vision transformers, however, they require sophisticated designs and massive datasets to work well. In this paper, we explore the idea of nesting basic local transformers on non-overlapping image blocks and aggregating them in a hierarchical way. We find that the block aggregation function plays a critical role in enabling cross-block non-local information communication. This observation leads us to design a simplified architecture that requires minor code changes upon the original vision transformer. The benefits of the proposed judiciously-selected design are threefold: (1)~NesT converges faster and requires much less training data to achieve good generalization on both ImageNet and small datasets like CIFAR; (2)~when extending our key ideas to image generation, NesT leads to a strong decoder that is 8$\times$ faster than previous transformer-based generators; and (3)~we show that decoupling the feature learning and abstraction processes via this nested hierarchy in our design enables constructing a novel method (named GradCAT) for visually interpreting the learned model.

poster session 1 @ red 3, poster session 11 @ red 3, oral session 1 @ red 3, poster session 1, poster session 11, oral session 1

Spatio-temporal representation learning is critical for video self-supervised representation. Recent approaches mainly use contrastive learning and pretext tasks. However, these approaches learn representation by discriminating sampled instances via feature similarity in the latent space while ignoring the intermediate state of the learned representations, which limits the overall performance. In this work, taking into account the degree of similarity of sampled instances as the intermediate state, we propose a novel pretext task - spatio-temporal overlap rate (STOR) prediction. It stems from the observation that humans are capable of discriminating the overlap rates of videos in space and time. This task encourages the model to discriminate the STOR of two generated samples to learn the representations. Moreover, we employ a joint optimization combining pretext tasks with contrastive learning to further enhance the spatio-temporal representation learning. We also study the mutual influence of each component in the proposed scheme. Extensive experiments demonstrate that our proposed STOR task can favor both contrastive learning and pretext tasks and the joint optimization scheme can significantly improve the spatio-temporal representation in video understanding. The code is available at https://github.com/Katou2/CSTP.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

The predefined artificially-balanced training classes in object recognition have limited capability in modeling real-world scenarios where objects are imbalanced-distributed with unknown classes. In this paper, we discuss a promising solution to the Open-set Long-Tailed Recognition (OLTR) task utilizing metric learning. Firstly, we propose a distribution-sensitive loss, which weighs more on the tail classes to decrease the intra-class distance in the feature space. Building upon these concentrated feature clusters, a local-density-based metric is introduced, called Localizing Unfamiliarity Near Acquaintance (LUNA), to measure the novelty of a testing sample. LUNA is flexible with different cluster sizes and is reliable on the cluster boundary by considering neighbors of different properties. Moreover, contrary to most of the existing works that alleviate the open-set detection as a simple binary decision, LUNA is a quantitative measurement with interpretable meanings. Our proposed method exceeds the state-of-the-art algorithm by 4-6% in the closed-set recognition accuracy and 4% in F-measure under the open-set on the public benchmark datasets, including our own newly introduced fine-grained OLTR dataset about marine species (MS-LT), which is the first naturally-distributed OLTR dataset revealing the genuine genetic relationships of the classes.

poster session 1 @ red 3, poster session 11 @ red 3, oral session 11 @ red 3, poster session 1, poster session 11, oral session 11

Open Set Recognition (OSR) has been an emerging topic. Besides recognizing predefined classes, the system needs to reject the unknowns. Prototype learning is a potential manner to handle the problem, as its ability to improve intra-class compactness of representations is much needed in discrimination between the known and the unknowns. In this work, we propose a novel Prototype Mining And Learning (PMAL) framework. It has a prototype mining mechanism before the phase of optimizing embedding space, explicitly considering two crucial properties, namely high-quality and diversity of the prototype set. Concretely, a set of high-quality candidates are firstly extracted from training samples based on data uncertainty learning, avoiding the interference from unexpected noise. Considering the multifarious appearance of objects even in a single category, a diversity-based strategy for prototype set filtering is proposed. Accordingly, the embedding space can be better optimized to discriminate therein the predefined classes and between known and unknowns. Extensive experiments verify the two good characteristics (i.e., high-quality and diversity) embraced in prototype mining, and show the remarkable performance of the proposed framework compared to state-of-the-arts.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Transformers have sprung up in the field of computer vision. In this work, we explore whether the core self-attention module in Transformer is the key to achieving excellent performance in image recognition. To this end, we build an attention-free network called sMLPNet based on the existing MLP-based vision models. Specifically, we replace the MLP module in the token-mixing step with a novel sparse MLP (sMLP) module. For 2D image tokens, sMLP applies 1D MLP along the axial directions and the parameters are shared among rows or columns. By sparse connection and weight sharing, sMLP module significantly reduces the number of model parameters and computational complexity, avoiding the common over-fitting problem that plagues the performance of MLP-like models. When only trained on the ImageNet-1K dataset, the proposed sMLPNet achieves 81.9% top-1 accuracy with only 24M parameters, which is much better than most CNNs and vision Transformers under the same model size constraint. When scaling up to 66M parameters, sMLPNet achieves 83.4% top-1 accuracy, which is on par with the state-of-the-art Swin Transformer. The success of sMLPNet suggests that the self-attention mechanism is not necessarily a silver bullet in computer vision. The code and models are publicly available at https://github.com/microsoft/SPACH.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Is it possible to complete Vision Transformer (ViT) pre-training without natural images and human-annotated labels? This question has become increasingly relevant in recent months because while current ViT pre-training tends to rely heavily on a large number of natural images and human-annotated labels, the recent use of natural images has resulted in problems related to privacy violation, inadequate fairness protection, and the need for labor-intensive annotations. In this paper, we experimentally verify that the results of formula-driven supervised learning (FDSL) framework are comparable with, and can even partially outperform, sophisticated self-supervised learning (SSL) methods like SimCLRv2 and MoCov2 without using any natural images in the pre-training phase. We also consider ways to reorganize FractalDB generation based on our tentative conclusion that there is room for configuration improvements in the iterated function system (IFS) parameter settings of such databases. Moreover, we show that while ViTs pre-trained without natural images produce visualizations that are somewhat different from ImageNet pre-trained ViTs, they can still interpret natural image datasets to a large extent. Finally, in experiments using the CIFAR-10 dataset, we show that our model achieved a performance rate of 97.8, which is comparable to the rate of 97.4 achieved with SimCLRv2 and 98.0 achieved with ImageNet.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

This paper proposes a representational model for image pairs such as consecutive video frames that are related by local pixel displacements, in the hope that the model may shed light on motion perception in primary visual cortex (V1). The model couples the following two components: (1) The vector representations of local contents of images and (2) The matrix representations of local pixel displacements caused by the relative motions between the agent and the objects in the 3D scene. When the image frame undergoes changes due to local pixel displacements, the vectors are multiplied by the matrices that represent the local displacements. Our experiments show that our model can learn Gabor-like filter pairs of quadrature phases. The profiles of the learned filters match those of simple cells in Macaque V1. Moreover, we demonstrate that the model can learn to infer local motions in either a supervised or unsupervised manner. With such a simple model, we achieve competitive results on optical flow estimation.

poster session 1 @ red 3, poster session 11 @ red 3, oral session 1 @ red 3, poster session 1, poster session 11, oral session 1

Knowledge Distillation (KD) is a widely-used technology to inherit information from cumbersome teacher models to compact student models, consequently realizing model compression and acceleration. Compared with image classification, object detection is a more complex task, and designing specific KD methods for object detection is non-trivial. In this work, we elaborately study the behaviour difference between the teacher and student detection models, and obtain two intriguing observations: First, the teacher and student rank their detected candidate boxes quite differently, which results in their precision discrepancy. Second, there is a considerable gap between the feature response differences and prediction differences between teacher and student, indicating that equally imitating all the feature maps of the teacher is the sub-optimal choice for improving the student's accuracy. Based on the two observations, we propose Rank Mimicking (RM) and Prediction-guided Feature Imitation (PFI) for distilling one-stage detectors, respectively. RM takes the rank of candidate boxes from teachers as a new form of knowledge to distill, which consistently outperforms the traditional soft label distillation. PFI attempts to correlate feature differences with prediction differences, making feature imitation directly help to improve the student's accuracy. On MS COCO and PASCAL VOC benchmarks, extensive experiments are conducted on various detectors with different backbones to validate the effectiveness of our method. Specifically, RetinaNet with ResNet50 achieves 40.4% mAP in MS COCO, which is 3.5% higher than its baseline, and also outperforms previous KD methods.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Few-shot object detection (FSOD) aims to detect objects using only a few examples. How to adapt state-of-the-art object detectors to the few-shot domain remains challenging. Object proposal is a key ingredient in modern object detectors. However, the quality of proposals generated for few-shot classes using existing methods is far worse than that of many-shot classes, e.g., missing boxes for few-shot classes due to misclassification or inaccurate spatial locations with respect to true objects. To address the noisy proposal problem, we propose a novel meta-learning based FSOD model by jointly optimizing the few-shot proposal generation and fine-grained few-shot proposal classification. To improve proposal generation for few-shot classes, we propose to learn a lightweight metric-learning based prototype matching network, instead of the conventional simple linear object/nonobject classifier, e.g., used in RPN. Our non-linear classifier with the feature fusion network could improve the discriminative prototype matching and the proposal recall for few-shot classes. To improve the fine-grained few-shot proposal classification, we propose a novel attentive feature alignment method to address the spatial misalignment between the noisy proposals and few-shot classes, thus improving the performance of few-shot object detection. Meanwhile we learn a separate Faster R-CNN detection head for many-shot base classes and show strong performance of maintaining base-classes knowledge. Our model achieves state-of-the-art performance on multiple FSOD benchmarks over most of the shots and metrics.

poster session 1 @ red 3, poster session 11 @ red 3, oral session 11 @ red 3, poster session 1, poster session 11, oral session 11

Vision transformers (ViTs) have been an alternative design paradigm to convolutional neural networks (CNNs). However, the training of ViTs is much harder than CNNs, as it is sensitive to the training parameters, such as learning rate, optimizer and warmup epoch. The reasons for training difficulty are empirically analysed in ~\cite{xiao2021early}, and the authors conjecture that the issue lies with the \textit{patchify-stem} of ViT models and propose that early convolutions help transformers see better. In this paper, we further investigate this problem and extend the above conclusion: only early convolutions do not help for stable training, but the scaled ReLU operation in the \textit{convolutional stem} (\textit{conv-stem}) matters. We verify, both theoretically and empirically, that scaled ReLU in \textit{conv-stem} not only improves training stabilization, but also increases the diversity of patch tokens, thus boosting peak performance with a large margin via adding few parameters and flops. In addition, extensive experiments are conducted to demonstrate that previous ViTs are far from being well trained, further showing that ViTs have great potential to be a better substitute of CNNs.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Object localization aims to generate a tight bounding box for the target object, which is a challenging problem that has been deeply studied in recent years. Since collecting bounding-box labels is time-consuming and laborious, many researchers focus on weakly supervised object localization (WSOL). As the recent appealing self-supervised learning technique shows its powerful function in visual tasks, in this paper, we take the early attempt to explore unsupervised object localization by self-supervision. Specifically, we adopt different geometric transformations to image and utilize their parameters as pseudo labels for self-supervised learning. Then, the class-agnostic activation map (CAAM) is used to highlight the target object potential regions. However, such attention maps merely focus on the most discriminative part of the objects, which will affect the quality of the predicted bounding box. Based on the motivation that the activation maps of different transformations of the same image should be equivariant, we further design a siamese network that encodes the paired images and propose a joint graph cluster partition mechanism in an unsupervised manner to enhance the object co-occurrent regions. To validate the effectiveness of the proposed method, extensive experiments are conducted on CUB-200-2011, Stanford Cars and FGVC-Aircraft datasets. Experimental results show that our method outperforms state-of-the-art methods using the same level of supervision, even outperforms some weakly-supervised methods.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

The recently proposed DEtection TRansformer (DETR) achieves promising performance for end-to-end object detection. However, it has relatively lower detection performance on small objects and suffers from slow convergence. This paper observed that DETR performs surprisingly well even on small objects when measuring Average Precision (AP) at decreased Intersection-over-Union (IoU) thresholds. Motivated by this observation, we propose a simple way to improve DETR by refining the coarse features and predicted locations. Specifically, we propose a novel Coarse-to-Fine (CF) decoder layer constituted of a coarse layer and a carefully designed fine layer. Within each CF decoder layer, the extracted local information (region of interest feature) is introduced into the flow of global context information from the coarse layer to refine and enrich the object query features via the fine layer. In the fine layer, the multi-scale information can be fully explored and exploited via the Adaptive Scale Fusion(ASF) module and Local Cross-Attention (LCA) module. The multi-scale information can also be enhanced by another proposed Transformer Enhanced FPN (TEF) module to further improve the performance. With our proposed framework (named CF-DETR), the localization accuracy of objects (especially for small objects) can be largely improved. As a byproduct, the slow convergence issue of DETR can also be addressed. The effectiveness of CF-DETR is validated via extensive experiments on the coco benchmark. CF-DETR achieves state-of-the-art performance among end-to-end detectors, e.g., achieving 47.8 AP using ResNet-50 with 36 epochs in the standard 3x training schedule.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Data in the real world tends to exhibit a long-tailed label distribution, which poses great challenges for the training of neural networks in visual recognition. Existing methods tackle this problem mainly from the perspective of data quantity, i.e., the number of samples in each class. To be specific, they pay more attention to tail classes, like applying larger adjustments to the logit. However, in the training process, the quantity and difficulty of data are two intertwined and equally crucial problems. For some tail classes, the features of their instances are distinct and discriminative, which can also bring satisfactory accuracy; for some head classes, although with sufficient samples, the high semantic similarity with other classes and lack of discriminative features will bring bad accuracy. Based on these observations, we propose Adaptive Logit Adjustment Loss (ALA Loss) to apply an adaptive adjusting term to the logit. The adaptive adjusting term is composed of two complementary factors: 1) quantity factor, which pays more attention to tail classes, and 2) difficulty factor, which adaptively pays more attention to hard instances in the training process. The difficulty factor can alleviate the over-optimization on tail yet easy instances and under-optimization on head yet hard instances. The synergy of the two factors can not only advance the performance on tail classes even further, but also promote the accuracy on head classes. Unlike previous logit adjusting methods that only concerned about data quantity, ALA Loss tackles the long-tailed problem from a more comprehensive, fine-grained and adaptive perspective. Extensive experimental results show that our method achieves the state-of-the-art performance on challenging recognition benchmarks, including ImageNet-LT, iNaturalist 2018, and Places-LT.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

The development of detection methods for oriented object detection remains a challenging task. A considerable obstacle is the wide variation in the shape (e.g., aspect ratio) of objects. Sample selection in general object detection has been widely studied as it plays a crucial role in the performance of the detection method and has achieved great progress. However, existing sample selection strategies still overlook some issues: (1) most of them ignore the object shape information; (2) they do not make a potential distinction between selected positive samples; and (3) some of them can only be applied to either anchor-free or anchor-based methods and cannot be used for both of them simultaneously. In this paper, we propose novel flexible shape-adaptive selection (SA-S) and shape-adaptive measurement (SA-M) strategies for oriented object detection, which comprise an SA-S strategy for sample selection and SA-M strategy for the quality estimation of positive samples. Specifically, the SA-S strategy dynamically selects samples according to the shape information and characteristics distribution of objects. The SA-M strategy measures the localization potential and adds quality information on the selected positive samples. The experimental results on both anchor-free and anchor-based baselines and four publicly available oriented datasets (DOTA, HRSC2016, UCAS-AOD, and ICDAR2015) demonstrate the effectiveness of the proposed method.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Pseudo bounding boxes from the self-training paradigm are inevitably noisy for semi-supervised object detection. To cope with that, a dual decoupling training framework is proposed in the present study, i.e. clean and noisy data decoupling, and classifcation and localization task decoupling. In the frst decoupling, two-level thresholds are used to categorize pseudo boxes into three groups, i.e. clean backgrounds, noisy foregrounds and clean foregrounds. With a specially designed noise-bypass head focusing on noisy data, backbone networks can extract coarse but diverse information; and meanwhile, an original head learns from clean samples for more precise predictions. In the second decoupling, we take advantage of the two-head structure for better evaluation of localization quality, thus the category label and location of a pseudo box can remain independent of each other during training. The approach of two-level thresholds is also applied to group pseudo boxes into three sections of different location accuracy. We outperform existing works by a large margin on VOC datasets, reaching 54.8 mAP(+1.8), and even up to 55.9 mAP(+1.5) by leveraging MS-COCO train2017 as extra unlabeled data. On MS-COCO benchmark, our method also achieves about 1.0 mAP improvements averaging across protocols compared with the prior state-of-the-art.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

We study the unsupervised representation learning for the semantic segmentation task. Different from previous works that aim at providing unsupervised pre-trained backbones for segmentation models which need further supervised fine-tune, here, we focus on providing representation that is only trained by unsupervised methods. This means models need to directly generate pixel-level, linearly separable semantic results. We first explore and present two factors that have significant effects on segmentation under the contrastive learning framework: 1) the difficulty and diversity of the positive contrastive pairs, 2) the balance of global and local features. With the intention of optimizing these factors, we propose the cycle-attention contrastive learning (CACL). CACL makes use of semantic continuity of video frames, adopting unsupervised cycle-consistent attention mechanism to implicitly conduct contrastive learning with difficult, global-local-balanced positive pixel pairs. Compared with baseline model MoCo-v2 and other unsupervised methods, CACL demonstrates consistently superior performance on PASCAL VOC (+4.5 mIoU) and Cityscapes (+4.5 mIoU) datasets.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Large-scale transformer-based methods are gaining popularity as a tool for predicting the properties of chemical compounds, which is of central importance to the drug discovery process. To accelerate their development and dissemination among the community, we are releasing HuggingMolecules -- an open-source library, with a simple and unified API, that provides the implementation of several state-of-the-art transformers for molecular property prediction. In addition, we add a comparison of these methods on several regression and classification datasets. HuggingMolecules package is available at: github.com/gmum/huggingmolecules.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Attention mechanism has been widely believed as the key to success of vision transformers (ViTs), since it provides a flexible and powerful way to model spatial relationships. However, is the attention mechanism truly an indispensable part of ViT? Can it be replaced by some other alternatives? To demystify the role of attention mechanism, we simplify it into an extremely simple case: ZERO FLOP and ZERO parameter. Concretely, we revisit the shift operation. It does not contain any parameter or arithmetic calculation. The only operation is to exchange a small portion of the channels between neighboring features. Based on this simple operation, we construct a new backbone network, namely ShiftViT, where the attention layers in ViT are substituted by shift operations. Surprisingly, ShiftViT works quite well in several mainstream tasks, e.g., classification, detection, and segmentation. The performance is on par with or even better than the strong baseline Swin Transformer. These results suggest that the attention mechanism might not be the vital factor that makes ViT successful. It can be even replaced by a zero-parameter operation. We should pay more attentions to the remaining parts of ViT in the future work. Code is available at github.com/microsoft/SPACH.

poster session 1 @ red 3, poster session 11 @ red 3, poster session 1, poster session 11

Given the incomplete knowledge of classes that exist in the world, Open-set Recognition (OSR) enables networks to identify and reject the unseen classes after training. This problem of breaking the common closed-set assumption is far from being solved. Recent studies focus on designing new losses, neural network encoding structures, and calibration methods to optimize a feature space for OSR relevant tasks. In this work, we make the first attempt to tackle OSR by searching the architecture of a Neural Network (NN) under the open-set assumption. In contrast to the prior arts, we develop a mechanism to both search the architecture of the network and train a network suitable for tackling OSR. Inspired by the compact abating probability (CAP) model, which is theoretically proven to reduce the open space risk, we regularize the searching space by VAE contrastive learning. To discover a more robust structure for OSR, we propose Pseudo Auxiliary Searching (PAS), in which we split a pretended set of know-unknown classes from the original training set in the searching phase, hence enabling the super-net to explore an effective architecture that can handle unseen classes in advance. We demonstrate the benefits of this learning pipeline on 5 OSR datasets, including MNIST, SVHN, CIFAR10, CIFARAdd10, and CIFARAdd50, where our approach outperforms prior state-of-the-art networks designed by humans.

poster session 1 @ red 3, poster session 11 @ red 3, oral session 1 @ red 3, poster session 1, poster session 11, oral session 1