Papers

Digital cameras transform sensor RAW readings into RGB images by means of their Image Signal Processor (ISP). Computational photography tasks such as image denoising and colour constancy are commonly performed in the RAW domain, in part due to the inherent hardware design, but also due to the appealing simplicity of noise statistics that result from the direct sensor readings. Despite this, the availability of RAW images is limited in comparison with the abundance and diversity of available RGB data. Recent approaches have attempted to bridge this gap by estimating the RGB to RAW mapping: handcrafted model-based methods that are interpretable and controllable usually require manual parameter fine-tuning, while end-to-end learnable neural networks require large amounts of training data, at times with complex training procedures, and generally lack interpretability and parametric control. Towards addressing these existing limitations, we present a novel hybrid model-based and data-driven ISP that builds on canonical ISP operations and is both learnable and interpretable. Our proposed invertible model, capable of bidirectional mapping between RAW and RGB domains, employs end-to-end learning of rich parameter representations, i.e. dictionaries, that are free from direct parametric supervision and additionally enable simple and plausible data augmentation. We evidence the value of our data generation process by extensive experiments under both RAW image reconstruction and RAW image denoising tasks, obtaining state-of-the-art performance in both. Additionally, we show that our ISP can learn meaningful mappings from few data samples, and that denoising models trained with our dictionary-based data augmentation are competitive despite having only few or zero ground-truth labels.

poster session 6 @ red 3, poster session 10 @ red 3, oral session 10 @ red 3, poster session 6, poster session 10, oral session 10

Pose-invariant face recognition is a practically useful but challenging task. This paper introduces a novel method to learn pose-invariant feature representation without normalizing profile faces to frontal ones or learning disentangled features. We first design a novel strategy to learn pose-invariant feature embeddings by distilling the angular knowledge of frontal faces extracted by teacher network to student network, which enables the handling of faces with large pose variations. In this way, the features of faces across variant poses can cluster compactly for the same person to create a pose-invariant face representation. Secondly, we propose a Pose-Adaptive Angular Distillation loss to mitigate the negative effect of uneven distribution of face poses in the training dataset to pay more attention to the samples with large pose variations. Extensive experiments on two challenging benchmarks (IJB-A and CFP-FP) show that our approach consistently outperforms the existing methods.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Conventional convolution neural networks (CNNs) trained on narrow Field-of-View (FoV) images are the state-of-the art approaches for object recognition tasks. Some methods proposed the adaptation of CNNs to ultra-wide FoV images by learning deformable kernels. However, they are limited by the Euclidean geometry and their accuracy degrades under strong distortions caused by fisheye projections. In this work, we demonstrate that learning the shape of convolution kernels in non-Euclidean spaces is better than existing deformable kernel methods. In particular, we propose a new approach that learns deformable kernel parameters (positions) in hyperbolic space. FisheyeHDK is a hybrid CNN architecture combining hyperbolic and Euclidean convolution layers for positions and features learning. First, we provide intuition of hyperbolic space for wide FoV images. Using synthetic distortion profiles, we demonstrate the effectiveness of our approach. We select two datasets - Cityscapes and BDD100K 2020 - of perspective images which we transform to fisheye equivalents at different scaling factors (analogue to focal lengths). Finally, we provide an experiment on data collected by a real fisheye camera. Validations and experiments show that our approach improves existing deformable kernel methods for CNN adaptation on fisheye images.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Recent neural rendering techniques have greatly benefited image-based modeling and relighting tasks. They provide a continuous, compact, and parallelable representation by modeling the plenoptic function as multilayer perceptrons (MLPs). However, vanilla MLPs suffer from spectral biases on multidimensional datasets. Recent rescues based on isotropic Fourier features mapping mitigate the problem but still fall short of handling heterogeneity across different dimensions, causing imbalanced regression and visual artifacts such as excessive blurs. We present an anisotropic random Fourier features (RFF) mapping scheme to tackle spectral biases. We first analyze the influence of bandwidth from a different perspective: we show that the optimal bandwidth exhibits strong correlations with the frequency spectrum of the training data across various dimensions. We then introduce an anisotropic feature mapping scheme with multiple bandwidths to model the multidimensional signal characteristics. We further propose an efficient bandwidth searching scheme through iterative golden-section search that can significantly reduce the training overload from polynomial time to logarithm. Our anisotropic scheme directly applies to neural surface light-field rendering and image-based relighting. Comprehensive experiments show that our scheme can more faithfully model lighting conditions and object features as well as preserve fine texture details and smooth view transitions even when angular and spatial samples are highly imbalanced.

poster session 6 @ red 3, poster session 10 @ red 3, oral session 10 @ red 3, poster session 6, poster session 10, oral session 10

Though deep learning-based object detection methods have achieved promising results on the conventional datasets, it is still challenging to locate objects from the low-quality images captured in adverse weather conditions. The existing methods either have difficulties in balancing the tasks of image enhancement and object detection, or often ignore the latent information beneficial for detection. To alleviate this problem, we propose a novel Image-Adaptive YOLO (IA-YOLO) framework, where each image can be adaptively enhanced for better detection performance. Specifically, a differentiable image processing (DIP) module is presented to take into account the adverse weather conditions for YOLO detector, whose parameters are predicted by a small convolutional neural network (CNN-PP). We learn CNN-PP and YOLOv3 jointly in an end-to-end fashion, which ensures that CNN-PP can learn an appropriate DIP to enhance the image for detection in a weakly supervised manner. Our proposed IA-YOLO approach can adaptively process images in both normal and adverse weather conditions. The experimental results are very encouraging, demonstrating the effectiveness of our proposed IA-YOLO method in both foggy and low-light scenarios. The source code can be found at https://github.com/wenyyu/Image-Adaptive-YOLO.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Recently, contrastive learning has largely advanced the progress of unsupervised visual representation learning. Pre-trained on ImageNet, some self-supervised algorithms reported higher transfer learning performance compared to fully-supervised methods, seeming to deliver the message that human labels hardly contribute to learning transferrable visual features. In this paper, we defend the usefulness of semantic labels but point out that fully-supervised and self-supervised methods are pursuing different kinds of features. To alleviate this issue, we present a new algorithm named Supervised Contrastive Adjustment in Neighborhood (SCAN) that maximally prevents the semantic guidance from damaging the appearance feature embedding. In a series of downstream tasks, SCAN achieves superior performance compared to previous fully-supervised and self-supervised methods, and sometimes the gain is significant. More importantly, our study reveals that semantic labels are useful in assisting self-supervised methods, opening a new direction for the community.

poster session 6 @ red 3, poster session 10 @ red 3, oral session 6 @ red 3, poster session 6, poster session 10, oral session 6

Few-shot learning (FSL) aims to classify images under low-data regimes, where the conventional pooled global feature is likely to lose useful local characteristics. Recent work has achieved promising performances by using deep descriptors. They generally take all deep descriptors from neural networks into consideration while ignoring that some of them are useless in classification due to their limited receptive field, e.g., task-irrelevant descriptors could be misleading and multiple aggregative descriptors from background clutter could even overwhelm the object's presence. In this paper, we argue that a Mutual Nearest Neighbor (MNN) relation should be established to explicitly select the query descriptors that are most relevant to each task and discard less relevant ones from aggregative clutters in FSL. Specifically, we propose Discriminative Mutual Nearest Neighbor Neural Network (DMN4) for FSL. Extensive experiments demonstrate that our method outperforms the existing state-of-the-arts on both fine-grained and generalized datasets.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

We propose a novel scene flow estimation approach to capture and infer 3D motions from point clouds. Estimating 3D motions for point clouds is challenging, since a point cloud is unordered and its density is significantly non-uniform. Such unstructured data poses difficulties in matching corresponding points between point clouds, leading to inaccurate flow estimation. We propose a novel architecture named Sparse Convolution-Transformer Network (SCTN) that equips the sparse convolution with the transformer. Specifically, by leveraging the sparse convolution, SCTN transfers irregular point cloud into locally consistent flow features for estimating spatially consistent motions within an object/local object part. We further propose to explicitly learn point relations using a point transformer module, different from exiting methods. We show that the learned relation-based contextual information is rich and helpful for matching corresponding points, benefiting scene flow estimation. In addition, a novel loss function is proposed to adaptively encourage flow consistency according to feature similarity. Extensive experiments demonstrate that our proposed approach achieves a new state of the art in scene flow estimation. Our approach achieves an error of 0.038 and 0.037 (EPE3D) on FlyingThings3D and KITTI Scene Flow respectively, which significantly outperforms previous methods by large margins.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Most metric learning techniques typically focus on sample embedding learning, while implicitly assume a homogeneous local neighborhood around each sample, based on the metrics used in training (\textit{e.g.,} hypersphere for Euclidean distance or unit hyperspherical crown for cosine distance). As real-world data often lies on a low-dimensional manifold curved in a high-dimensional space, it is unlikely that everywhere of the manifold shares the same local structures in the input space. Besides, considering the non-linearity of neural networks, the local structure in the output embedding space may not be homogeneous as assumed. Therefore, representing each sample simply with its embedding while ignoring its individual neighborhood structure would have limitations in Embedding-Based Retrieval (EBR). By exploiting the heterogeneity of local structures in the embedding space, we propose a Neighborhood-Adaptive Structure Augmented metric learning framework (NASA), where the neighborhood structure is realized as a \textit{structure embedding}, and learned along with the sample embedding in a self-supervised manner. In this way, without any modifications, most indexing techniques can be used to support large-scale EBR with NASA embeddings. Experiments on six standard benchmarks with two kinds of embeddings, \textit{i.e.,} binary embeddings and real-valued embeddings, show that our method significantly improves and outperforms the state-of-the-art methods.

poster session 6 @ red 3, poster session 10 @ red 3, oral session 6 @ red 3, poster session 6, poster session 10, oral session 6

Training effective Generative Adversarial Networks (GANs) requires large amounts of training data, without which the trained models are usually sub-optimal with discriminator over-fitting. Several prior studies address this issue by expanding the distribution of the limited training data via massive and hand-crafted data augmentation. We handle data-limited image generation from a very different perspective. Specifically, we design GenCo, a Generative Co-training network that mitigates the discriminator over-fitting issue by introducing multiple complementary discriminators that provide diverse supervision from multiple distinctive views in training. We instantiate the idea of GenCo in two ways. The first way is Weight-Discrepancy Co-training (WeCo) which co-trains multiple distinctive discriminators by diversifying their parameters. The second way is Data-Discrepancy Co-training (DaCo) which achieves co-training by feeding discriminators with different views of the input images (e.g., different frequency components of the input images). Extensive experiments over multiple benchmarks show that GenCo achieves superior generation with limited training data. In addition, GenCo also complements the augmentation approach with consistent and clear performance gains when combined.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Image-guided depth completion aims to generate dense depth maps with sparse depth measurements and corresponding RGB images. Currently, spatial propagation networks (SPNs) are the most popular affinity-based methods in depth completion, but they still suffer from the representation limitation of the fixed affinity and the over smoothing during iterations. Our solution is to estimate independent affinity matrices in each SPN iteration, but it is over-parameterized and heavy calculation.This paper introduces an efficient model that learns the affinity among neighboring pixels with an attention-based, dynamic approach. Specifically, the Dynamic Spatial Propagation Network (DySPN) we proposed makes use of a non-linear propagation model (NLPM). It decouples the neighborhood into parts regarding to different distances and recursively generates independent attention maps to refine these parts into adaptive affinity matrices. Furthermore, we adopt a diffusion suppression (DS) operation so that the model converges at an early stage to prevent over-smoothing of dense depth. Finally, in order to decrease the computational cost required, we also introduce three variations that reduce the amount of neighbors and attentions needed while still retaining similar accuracy. In practice, our method requires less iteration to match the performance of other SPNs and yields better results overall. DySPN outperforms other state-of-the-art (SoTA) methods on KITTI Depth Completion (DC) evaluation by the time of submission and is able to yield SoTA performance in NYU Depth v2 dataset as well.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Image composition targets at inserting a foreground object into a background image. Most previous image composition methods focus on adjusting the foreground to make it compatible with background while ignoring the shadow effect of foreground on the background. In this work, we focus on generating plausible shadow for the foreground object in the composite image. First, we contribute a real-world shadow generation dataset DESOBA by generating synthetic composite images based on paired real images and deshadowed images. Then, we propose a novel shadow generation network SGRNet, which consists of a shadow mask prediction stage and a shadow filling stage. In the shadow mask prediction stage, foreground and background information are thoroughly interacted to generate foreground shadow mask. In the shadow filling stage, shadow parameters are predicted to fill the shadow area. Extensive experiments on our DESOBA dataset and real composite images demonstrate the effectiveness of our proposed method. Our dataset and code are available at https://github.com/bcmi/Object-Shadow-Generation-Dataset-DESOBA.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Recent advances in supervised deep learning methods are enabling remote measurements of photoplethysmography-based physiological signals using facial videos. The performance of these supervised methods, however, are dependent on the availability of large labelled data. Contrastive learning as a self-supervised method has recently achieved state-of-the-art performances in learning representative data features by maximising mutual information between different augmented views. However, existing data augmentation techniques for contrastive learning are not designed to learn physiological signals from videos and often fail when there are complicated noise and subtle and periodic colour/shape variations between video frames. To address these problems, we present a novel self-supervised spatiotemporal learning framework for remote physiological signal representation learning, where there is a lack of labelled training data. Firstly, we propose a landmark-based spatial augmentation that splits the face into several informative parts based on the Shafer’s dichromatic reflection model to characterise subtle skin colour fluctuations. We also formulate a sparsity-based temporal augmentation exploiting Nyquist–Shannon sampling theorem to effectively capture periodic temporal changes by modelling physiological signal features. Furthermore, we introduce a constrained spatiotemporal loss that generates pseudo-labels for augmented video clips. It is used to regulate the training process and handle complicated noise. We evaluated our framework on 3 public datasets and demonstrated superior performances than other self-supervised methods and achieved competitive accuracy compared to the state-of-the-art supervised methods.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

We attempt to train deep neural networks for classification without using any labeled data. Existing unsupervised methods, though mine useful clusters or features, require some annotated samples to facilitate the final task-specific predictions. This defeats the true purpose of unsupervised learning and hence we envisage a paradigm of `true' self-supervision, where absolutely no annotated instances are used for training a classifier. The proposed method first pretrains a deep network through self-supervision and performs clustering on the learned features. A classifier layer is then appended to the self-supervised network and is trained by matching the distribution of the predictions to that of a predefined prior. This approach leverages the distribution of labels for supervisory signals and consequently, no image-label pair is needed. Experiments reveal that the method works on major nominal as well as ordinal classification datasets and delivers significant performance.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Recently, automated surveillance cameras can change a visible sensor and a thermal sensor for all-day operation. However, existing single-modal pedestrian detectors mainly focus on detecting pedestrians in only one specific modality (i.e., visible or thermal), so they cannot effectively cope with other modal inputs. In addition, recent multispectral pedestrian detectors have shown remarkable performance by adopting multispectral modalities, but they also have limitations in practical applications (e.g., different Field-of-View (FoV) and frame rate). In this paper, we introduce a versatile pedestrian detector that shows robust detection performance in any single modality. We propose a multisensory-matching contrastive loss to reduce the difference between the visual representation of pedestrians in the visible and the thermal modalities. Moreover, to make the proposed method perform robust detection on a single modality, we design a Multispectral Recalling (MSR) Memory. The MSR Memory enhances the visual representation of the single modal features by recalling that of the multispectral modalities. To guide the MSR Memory to store the contexts of the multispectral modalities, we introduce a multispectral recalling loss. It enables the pedestrian detector to encode more discriminative features with a single input modality. We would like to insist that our method is a step forward detector that can be applied to a variety of real-world applications. The comprehensive experimental results verify the effectiveness of the proposed method.

poster session 6 @ red 3, poster session 10 @ red 3, oral session 6 @ red 3, poster session 6, poster session 10, oral session 6

Feature Pyramid Network (FPN) has been widely adopted to exploit multi-scale features for scale variation in object detection. However, intrinsic defects in most of the current methods with FPN make it difficult to adapt to the feature of different geometric objects. To address this issue, we introduce geometric prior into FPN to obtain more discriminative features. In this paper, we propose Geometry-aware Feature Pyramid Network (GaFPN), which consists of Geometry-aware Mapping Module, Geometry-aware Predictor Head.Geometry-aware Mapping Module is proposed to make full use of all pyramid features to obtain better proposal features by the weight-generation subnetwork. The weights generation subnetwork generates fusion weight for each layer proposal features by using the geometric information of the proposal. Geometry-aware Predictor Head introduces geometric prior into predictor head by the embedding generation net-work to strengthen feature representation for classification and regression. Our GaFPN can be easily extended to other two-stage object detectors with feature pyramid and applied to instance segmentation task. The proposed GaFPN significantly improves detection performance compared to baseline detectors with ResNet-50-FPN: +1.9, +2.0, +1.7, +1.3, +0.8 points Average Precision (AP) on Faster-RCNN, Cascade R-CNN, Dynamic R-CNN, SABL, and AugFPN respectively on MS COCO dataset.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

This paper introduces Ranking Info Noise Contrastive Estimation (RINCE), a new member in the family of InfoNCE losses that preserves a ranked ordering of positive samples. In contrast to the standard InfoNCE loss, which requires a strict binary separation of the training pairs into similar and dissimilar samples, RINCE can exploit information about a similarity ranking for learning a corresponding embedding space. We show that the proposed loss function learns favorable embeddings compared to the standard InfoNCE whenever at least noisy ranking information can be obtained or when the definition of positives and negatives is blurry. We demonstrate this for a supervised classification task with additional superclass labels and noisy similarity scores. Furthermore, we show that RINCE can also be applied to unsupervised training with experiments on unsupervised representation learning from videos. In particular, the embedding yields higher classification accuracy, retrieval rates and performs better on out-of-distribution detection than the standard InfoNCE loss.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Monocular 3D object detection aims to localize 3D bounding boxes in an input single 2D image. It is a highly challenging problem and remains open, especially when no extra information (e.g., depth, lidar and/or multi-frames) can be leveraged in training and/or inference. This paper proposes a simple yet effective formulation for monocular 3D object detection without exploiting any extra information. It presents the MonoCon method which learns Monocular Contexts, as auxiliary tasks in training, to help monocular 3D object detection. The key idea is that with the annotated 3D bounding boxes of objects in an image, there is a rich set of well-posed projected 2D supervision signals available in training, such as the projected corner keypoints and their associated offset vectors with respect to the center of 2D bounding box, which should be exploited as auxiliary tasks in training. The proposed MonoCon is motivated by the Cram\`er–Wold theorem in measure theory at a high level. In implementation, it utilizes a very simple end-to-end design to justify the effectiveness of learning auxiliary monocular contexts, which consists of three components: a Deep Neural Network (DNN) based feature backbone, a number of regression head branches for learning the essential parameters used in the 3D bounding box prediction, and a number of regression head branches for learning auxiliary contexts. After training, the auxiliary context regression branches are discarded for better inference efficiency. In experiments, the proposed MonoCon is tested in the KITTI benchmark (car, pedestrian and cyclist). It outperforms all prior arts in the leaderboard on the car category and obtains comparable performance on pedestrian and cyclist in terms of accuracy. Thanks to the simple design, the proposed MonoCon method obtains the fastest inference speed with 38.7 fps in comparisons.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

We propose a new object-centric framework for learning-based stereo 3D object detection. Previous studies build scene-centric representations that do not consider the significant variation among outdoor instances and thus lack the flexibility and functionalities that an instance-level model can offer. We build such an instance-level model by formulating and tackling a local update problem, i.e., how to predict a refined update given an initial 3D cuboid guess. We demonstrate how solving this problem can complement scene-centric approaches in (i) building a coarse-to-fine multi-resolution system, (ii) performing model-agnostic object location refinement, and (iii) conducting stereo 3D tracking-by-detection. Extensive experiments demonstrate the effectiveness of our approach, which achieves state-of-the-art performance on the KITTI benchmark. Code and pre-trained models are available at https://github.com/Nicholasli1995/SNVC.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Two major challenges of 3D LiDAR Panoptic Segmentation (PS) are that point clouds of an object are surface-aggregated and thus hard to model the long-range dependency especially for large instances, and that objects are too close to separate each other. Recent literature addresses these problems by time-consuming grouping processes such as dual-clustering, mean-shift offsets and etc., or by bird-eye-view (BEV) dense centroid representation that downplays geometry. However, the long-range geometry relationship has not been sufficiently modeled by local feature learning from the above methods. To this end, we present SCAN, a novel sparse cross-scale attention network to first align multi-scale sparse features with global voxel-encoded attention to capture the long-range relationship of instance context, which is able to boost the regression accuracy of the over-segmented large objects. For the surface-aggregated points, SCAN adopts a novel sparse class-agnostic representation of instance centroids, which can not only maintain the sparsity of aligned features to solve the under-segmentation on small objects, but also reduce the computation amount of the network through sparse convolution. Our method outperforms previous methods by a large margin in the SemanticKITTI dataset for the challenging 3D PS task, achieving 1st place with a real-time inference speed.

poster session 6 @ red 3, poster session 10 @ red 3, oral session 10 @ red 3, poster session 6, poster session 10, oral session 10

Orthogonality regularization has proven effective in improving the precision, convergence speed and the training stability of CNNs. Here, we propose a novel Orthogonal Dictionary Convolution Strategy (ODCS) on CNNs to improve orthogonality effect by optimizing the network architecture and changing the regularized object. Specifically, we remove the nonlinear layer in typical convolution block “Conv(BN) + Nonlinear + Pointwise Conv(BN)”, and only impose orthogonal regularization on the front Conv. The structure, “Conv(BN) + Pointwise Conv(BN)”, is then equivalent to a pair of dictionary and encoding, defined in sparse dictionary learning. Thanks to the exact and efficient representation of signal with dictionaries in low-dimensional projections, our strategy could reduce the superfluous information in dictionary Conv kernels. Meanwhile, the proposed strategy relieves the too strict orthogonality regularization in training, which makes hyper-parameters tuning of model to be more flexible. In addition, our ODCS can modify the state-of-the-art models easily without any extra consumption in inference phase. We evaluate it on a variety of CNNs in small-scale (CIFAR), large-scale (ImageNet) and fine-grained (CUB-200-2011) image classification tasks, respectively. The experimental results show that our method achieve a stable and superior improvement.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Recent techniques to solve photorealistic style transfer within deep convolutional neural networks (CNNs) generally require intensive training from large-scale datasets, thus having limited applicability and poor generalization ability to unseen images or styles. To overcome this, we propose a novel framework, dubbed Deep Translation Prior (DTP), to accomplish photorealistic style transfer through test-time training on given input image pair with untrained networks, which learns an image pair-specific translation prior and thus yields better performance and generalization. Tailored for such test-time training for style transfer, we present novel network architectures, with two sub-modules of correspondence and generation modules, and loss functions consisting of contrastive content, style, and cycle consistency losses. Our framework does not require offline training phase for style transfer, which has been one of the main challenges in existing methods, but the networks are to be solely learned during test time. Experimental results prove that our framework has a better generalization ability to unseen image pairs and even outperforms the state-of-the-art methods.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

In object re-identification (ReID), the development of deep learning techniques often involves model updates and deployment. It is unbearable to re-embedding and re-index with the system suspended when deploying new models. Therefore, backward-compatible representation is proposed to enable ``new'' features to be compared with ``old'' features directly, which means that the database is active when there are both ``new'' and ``old'' features in it. Thus we can scroll-refresh the database or even do nothing on the database to update. The existing backward-compatible methods either require a strong overlap between old and new training data or simply conduct constraints at the instance level. Thus they are difficult in handling complicated cluster structures and are limited in eliminating the impact of outliers in old embeddings, resulting in a risk of damaging the discriminative capability of new features. In this work, we propose a Neighborhood Consensus Contrastive Learning (NCCL) method. With no assumptions about the new training data, we estimate the sub-cluster structures of old embeddings. A new embedding is constrained with multiple old embeddings in both embedding space and discrimination space at the sub-class level. The effect of outliers diminished, as the multiple samples serve as ``mean teachers''. Besides, we also propose a scheme to filter the old embeddings with low credibility, further improving the compatibility robustness. Our method ensures backward compatibility without impairing the accuracy of the new model. And it can even improve the new model's accuracy in most scenarios.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10

Since people know the important factors that affect the control based on the driving situation, they can drive safely even in diverse driving environment. To mimic these behaviors, we propose a two-stage representation learning model that initially splits the latent features as domain-specific features, which contain information in a specific domain, and domain-general features, which are consistent across all domains. Subsequently, the dynamic-object features, which contain information of dynamic objects, are disentangled from latent features using mutual information estimator. In this study, the problem in behavior cloning is divided into several domain-specific subspaces, with experts becoming specialized on each domain-specific policy. The proposed mixture of domain-specific experts (MoDE) model predicts the final control values through the cooperation of experts using a gating function. The domain-specific features are used to calculate the importance weight of the domain-specific experts, and the disentangled domain-general and dynamic-object features are applied in estimating the control values. To validate the proposed MoDE model, we conducted several experiments and achieved a higher success rate on the CARLA benchmarks under several conditions and tasks than state-of-the-art approaches.

poster session 6 @ red 3, poster session 10 @ red 3, poster session 6, poster session 10