Depth estimation is a critical technology in autonomous driving, and multi-camera systems are often used to achieve a 360$^\circ$ perception. These 360$^\circ$ camera sets often have limited or low-quality overlap regions, making multi-view stereo methods infeasible for the entire image. Alternatively, monocular methods may not produce consistent cross-view predictions. To address these issues, we propose the Stereo Guided Depth Estimation (SGDE) method, which enhances depth estimation of the full image by explicitly utilizing multi-view stereo results on the overlap. We suggest building virtual pinhole cameras to resolve the distortion problem of fisheye cameras and unify the processing for the two types of 360$^\circ$ cameras. For handling the varying noise on camera poses caused by unstable movement, the approach employs a self-calibration method to obtain highly accurate relative poses of the adjacent cameras with minor overlap. These enable the use of robust stereo methods to obtain high-quality depth prior in the overlap region. This prior serves not only as an additional input but also as pseudo-labels that enhance the accuracy of depth estimation methods and improve cross-view prediction consistency. The effectiveness of SGDE is evaluated on one fisheye camera dataset, Synthetic Urban, and two pinhole camera datasets, DDAD and nuScenes. Our experiments demonstrate that SGDE is effective for both supervised and self-supervised depth estimation, and highlight the potential of our method for advancing downstream autonomous driving technologies, such as 3D object detection and occupancy prediction.
Monocular depth estimation from RGB images plays a pivotal role in 3D vision. However, its accuracy can deteriorate in challenging environments such as nighttime or adverse weather conditions. While long-wave infrared cameras offer stable imaging in such challenging conditions, they are inherently low-resolution, lacking rich texture and semantics as delivered by the RGB image. Current methods focus solely on a single modality due to the difficulties to identify and integrate faithful depth cues from both sources. To address these issues, this paper presents a novel approach that identifies and integrates dominant cross-modality depth features with a learning-based framework. Concretely, we independently compute the coarse depth maps with separate networks by fully utilizing the individual depth cues from each modality. As the advantageous depth spreads across both modalities, we propose a novel confidence loss steering a confidence predictor network to yield a confidence map specifying latent potential depth areas. With the resulting confidence map, we propose a multi-modal fusion network that fuses the final depth in an end-to-end manner. Harnessing the proposed pipeline, our method demonstrates the ability of robust depth estimation in a variety of difficult scenarios. Experimental results on the challenging MS$^2$ and ViViD++ datasets demonstrate the effectiveness and robustness of our method.
Visual SLAM (Simultaneous Localization and Mapping) based on planar features has found widespread applications in fields such as environmental structure perception and augmented reality. However, current research faces challenges in accurately localizing and mapping in planar ambiguous scenes, primarily due to the poor accuracy of the employed planar features and data association methods. In this paper, we propose a visual SLAM system based on planar features designed for planar ambiguous scenes, encompassing planar processing, data association, and multi-constraint factor graph optimization. We introduce a planar processing strategy that integrates semantic information with planar features, extracting the edges and vertices of planes to be utilized in tasks such as plane selection, data association, and pose optimization. Next, we present an integrated data association strategy that combines plane parameters, semantic information, projection IoU (Intersection over Union), and non-parametric tests, achieving accurate and robust plane data association in planar ambiguous scenes. Finally, we design a set of multi-constraint factor graphs for camera pose optimization. Qualitative and quantitative experiments conducted on publicly available datasets demonstrate that our proposed system competes effectively in both accuracy and robustness in terms of map construction and camera localization compared to state-of-the-art methods.
Deep learning systems are prone to catastrophic forgetting when learning from a sequence of tasks, where old data from experienced tasks is unavailable when learning from a new task. To mitigate the problem, a line of methods propose to replay the data of experienced tasks when learning new tasks. These methods usually adopt an extra memory to store the data for replay. However, it is not expected in practice considering the memory constraint or data privacy issue. As a replacement, data-free data replay methods are proposed by inverting samples from the classification model. Though achieving good results, these methods still suffer from the inconsistency of the inverted and real training data, which is neglected in the inversion stage in recent works. To that effect, we propose to measure the data consistency quantitatively by some simplification and assumptions. Using the measurement, we analyze existing techniques for inverting samples and get some insightful information that inspires a novel loss function to reduce the inconsistency. Specifically, the loss minimizes the KL divergence of the distributions of inverted and real data under the tied multivariate Gaussian assumption, which is easy to implement in continual learning. In addition, we observe that the norms of old class weights turn to decrease continually as learning progresses. We thus analyze the underlying reasons and propose a simple regularization term to balance the class weights so that the samples of old classes are more distinguishable. To conclude, we propose the Consistency enhanced data replay with debiased classifier for Class Incremental Learning (CCIL). Extensive experiments on CIFAR-100, Tiny-ImageNet, and ImageNet100 show consistently improved performance of CCIL compared to previous approaches.
Recent works have studied implicit biases in deep learning, especially the behavior of last-layer features and classifier weights. However, they usually need to simplify the intermediate dynamics under gradient flow or gradient descent due to the intractability of loss functions and model architectures. In this paper, we introduce the unhinged loss, a concise loss function, that offers more mathematical opportunities to analyze the closed-form dynamics while requiring as few simplifications or assumptions as possible. The unhinged loss allows for considering more practical techniques, such as time-vary learning rates and feature normalization. Based on the layer-peeled model that views last-layer features as free optimization variables, we conduct a thorough analysis in the unconstrained, regularized, and spherical constrained cases, as well as the case where the neural tangent kernel remains invariant. To bridge the performance of the unhinged loss to that of Cross-Entropy (CE), we investigate the scenario of fixing classifier weights with a specific structure, (e.g., a simplex equiangular tight frame). Our analysis shows that these dynamics converge exponentially fast to a solution depending on the initialization of features and classifier weights. These theoretical results not only offer valuable insights, including explicit feature regularization and rescaled learning rates for enhancing practical training with the unhinged loss, but also extend their applicability to other loss functions. Finally, we empirically demonstrate these theoretical results and insights through extensive experiments.
Reconstructing hand-held objects from a single RGB image without known 3D object templates, category prior, or depth information is a vital yet challenging problem in computer vision. In contrast to prior works that utilize deterministic modeling paradigms, which make it hard to account for the uncertainties introduced by hand- and self-occlusion, we employ a probabilistic point cloud denoising diffusion model to tackle the above challenge. In this work, we present Hand-Aware Conditional Diffusion for monocular hand-held object reconstruction (HACD), modeling the hand-object interaction in two aspects. First, we introduce hand-aware conditioning to model hand-object interaction from both semantic and geometric perspectives. Specifically, a unified hand-object semantic embedding compensates for the 2D local feature deficiency induced by hand occlusion, and a hand articulation embedding further encodes the relationship between object vertices and hand joints. Second, we propose a hand-constrained centroid fixing scheme, which utilizes hand vertices priors to restrict the centroid deviation of partially denoised point cloud during diffusion and reverse process. Removing the centroid bias interference allows the diffusion models to focus on the reconstruction of shape, thus enhancing the stability and precision of local feature projection. Experiments on the synthetic ObMan dataset and two real-world datasets, HO3D and MOW, demonstrate our approach surpasses all existing methods by a large margin.
In this paper, we present ShapeMaker, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonicalize the object by extracting point-wise affine-invariant features, disentangling inherent structure of the object with its pose and size. These learned features are then leveraged to predict semantically consistent part segmentation and corresponding part centers. Next, our lightweight retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. The key insight of ShapeMaker is the simultaneous training of the four highly-associated processes: canonicalization, segmentation, retrieval, and deformation, leveraging cross-task consistency losses for mutual supervision. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that ShapeMaker surpasses competitors by a large margin. Codes will be released soon.
Offline reinforcement learning suffers from the out-of-distribution issue and extrapolation error. Most policy constraint methods regularize the density of the trained policy towards the behavior policy, which is too restrictive in most cases. We propose Supported Trust Region optimization (STR) which performs trust region policy optimization with the policy constrained within the support of the behavior policy, enjoying the less restrictive support constraint. We show that, when assuming no approximation and sampling error, STR guarantees strict policy improvement until convergence to the optimal support-constrained policy in the dataset. Further with both errors incorporated, STR still guarantees safe policy improvement for each step. Empirical results validate the theory of STR and demonstrate its state-of-the-art performance on MuJoCo locomotion domains and much more challenging AntMaze domains.
Learning generalized representations from limited training samples is crucial for applying deep neural networks in low-resource scenarios. Recently, methods based on Contrastive Language-Image Pre-training (CLIP) have exhibited promising performance in few-shot adaptation tasks. To avoid catastrophic forgetting and overfitting caused by few-shot fine-tuning, existing works usually freeze the parameters of CLIP pre-trained on large-scale datasets, overlooking the possibility that some parameters might not be suitable for downstream tasks. To this end, we revisit CLIP's visual encoder with a specific focus on its distinctive attention pooling layer, which performs a spatial weighted-sum of the dense feature maps. Given that dense feature maps contain meaningful semantic information, and different semantics hold varying importance for diverse downstream tasks (such as prioritizing semantics like ears and eyes in pet classification tasks rather than side mirrors), using the same weighted-sum operation for dense features across different few-shot tasks might not be appropriate. Hence, we propose fine-tuning the parameters of the attention pooling layer during the training process to encourage the model to focus on task-specific semantics. In the inference process, we perform residual blending between the features pooled by the fine-tuned and the original attention pooling layers to incorporate both the few-shot knowledge and the pre-trained CLIP's prior knowledge. We term this method as Semantic-Aware FinE-tuning (SAFE). SAFE is effective in enhancing the conventional few-shot CLIP and is compatible with the existing adapter approach (termed SAFE-A).