A Dataset And Benchmark Of Underwater Object Detection For Robot Picking

Underwater object detection for robot picking has attracted a lot of interest. However, it is still an unsolved problem due to several challenges. We take steps towards making it more realistic by addressing the following challenges. Firstly, the currently available datasets basically lack the test set annotations, causing researchers must compare their method with other SOTAs on a self-divided test set (from the training set). Training other methods lead to an increase in workload and different researchers divide different datasets, resulting there is no unified benchmark to compare the performance of different algorithms. Secondly, these datasets also have other shortcomings, e.g., too many similar images or incomplete labels. Towards these challenges we introduce a dataset, Detecting Underwater Objects (DUO), and a corresponding benchmark, based on the collection and re-annotation of all relevant datasets. DUO contains a collection of diverse underwater images with more rational annotations. The corresponding benchmark provides indicators of both efficiency and accuracy of SOTAs (under the MMDtection framework) for academic research and industrial applications, where JETSON AGX XAVIER is used to assess detector speed to simulate the robot-embedded environment.

Delving into Localization Errors for Monocular 3D Object Detection

Estimating 3D bounding boxes from monocular images is an essential component in autonomous driving, while accurate 3D object detection from this kind of data is very challenging. In this work, by intensive diagnosis experiments, we quantify the impact introduced by each sub-task and found the `localization error' is the vital factor in restricting monocular 3D detection. Besides, we also investigate the underlying reasons behind localization errors, analyze the issues they might bring, and propose three strategies. First, we revisit the misalignment between the center of the 2D bounding box and the projected center of the 3D object, which is a vital factor leading to low localization accuracy. Second, we observe that accurately localizing distant objects with existing technologies is almost impossible, while those samples will mislead the learned network. To this end, we propose to remove such samples from the training set for improving the overall performance of the detector. Lastly, we also propose a novel 3D IoU oriented loss for the size estimation of the object, which is not affected by `localization error'. We conduct extensive experiments on the KITTI dataset, where the proposed method achieves real-time detection and outperforms previous methods by a large margin. The code will be made available at: https://github.com/xinzhuma/monodle.

Learning Scene Structure Guidance via Cross-Task Knowledge Transfer for Single Depth Super-Resolution

Existing color-guided depth super-resolution (DSR) approaches require paired RGB-D data as training samples where the RGB image is used as structural guidance to recover the degraded depth map due to their geometrical similarity. However, the paired data may be limited or expensive to be collected in actual testing environment. Therefore, we explore for the first time to learn the cross-modality knowledge at training stage, where both RGB and depth modalities are available, but test on the target dataset, where only single depth modality exists. Our key idea is to distill the knowledge of scene structural guidance from RGB modality to the single DSR task without changing its network architecture. Specifically, we construct an auxiliary depth estimation (DE) task that takes an RGB image as input to estimate a depth map, and train both DSR task and DE task collaboratively to boost the performance of DSR. Upon this, a cross-task interaction module is proposed to realize bilateral cross task knowledge transfer. First, we design a cross-task distillation scheme that encourages DSR and DE networks to learn from each other in a teacher-student role-exchanging fashion. Then, we advance a structure prediction (SP) task that provides extra structure regularization to help both DSR and DE networks learn more informative structure representations for depth recovery. Extensive experiments demonstrate that our scheme achieves superior performance in comparison with other DSR methods.

A Unified Joint Maximum Mean Discrepancy for Domain Adaptation

Domain adaptation has received a lot of attention in recent years, and many algorithms have been proposed with impressive progress. However, it is still not fully explored concerning the joint probability distribution (P(X, Y)) distance for this problem, since its empirical estimation derived from the maximum mean discrepancy (joint maximum mean discrepancy, JMMD) will involve complex tensor-product operator that is hard to manipulate. To solve this issue, this paper theoretically derives a unified form of JMMD that is easy to optimize, and proves that the marginal, class conditional and weighted class conditional probability distribution distances are our special cases with different label kernels, among which the weighted class conditional one not only can realize feature alignment across domains in the category level, but also deal with imbalance dataset using the class prior probabilities. From the revealed unified JMMD, we illustrate that JMMD degrades the feature-label dependence (discriminability) that benefits to classification, and it is sensitive to the label distribution shift when the label kernel is the weighted class conditional one. Therefore, we leverage Hilbert Schmidt independence criterion and propose a novel MMD matrix to promote the dependence, and devise a novel label kernel that is robust to label distribution shift. Finally, we conduct extensive experiments on several cross-domain datasets to demonstrate the validity and effectiveness of the revealed theoretical results.

Full Matching on Low Resolution for Disparity Estimation

A Multistage Full Matching disparity estimation scheme (MFM) is proposed in this work. We demonstrate that decouple all similarity scores directly from the low-resolution 4D volume step by step instead of estimating low-resolution 3D cost volume through focusing on optimizing the low-resolution 4D volume iteratively leads to more accurate disparity. To this end, we first propose to decompose the full matching task into multiple stages of the cost aggregation module. Specifically, we decompose the high-resolution predicted results into multiple groups, and every stage of the newly designed cost aggregation module learns only to estimate the results for a group of points. This alleviates the problem of feature internal competitive when learning similarity scores of all candidates from one low-resolution 4D volume output from one stage. Then, we propose the strategy of \emph{Stages Mutual Aid}, which takes advantage of the relationship of multiple stages to boost similarity scores estimation of each stage, to solve the unbalanced prediction of multiple stages caused by serial multistage framework. Experiment results demonstrate that the proposed method achieves more accurate disparity estimation results and outperforms state-of-the-art methods on Scene Flow, KITTI 2012 and KITTI 2015 datasets.

Direct Depth Learning Network for Stereo Matching

Being a crucial task of autonomous driving, Stereo matching has made great progress in recent years. Existing stereo matching methods estimate disparity instead of depth. They treat the disparity errors as the evaluation metric of the depth estimation errors, since the depth can be calculated from the disparity according to the triangulation principle. However, we find that the error of the depth depends not only on the error of the disparity but also on the depth range of the points. Therefore, even if the disparity error is low, the depth error is still large, especially for the distant points. In this paper, a novel Direct Depth Learning Network (DDL-Net) is designed for stereo matching. DDL-Net consists of two stages: the Coarse Depth Estimation stage and the Adaptive-Grained Depth Refinement stage, which are all supervised by depth instead of disparity. Specifically, Coarse Depth Estimation stage uniformly samples the matching candidates according to depth range to construct cost volume and output coarse depth. Adaptive-Grained Depth Refinement stage performs further matching near the coarse depth to correct the imprecise matching and wrong matching. To make the Adaptive-Grained Depth Refinement stage robust to the coarse depth and adaptive to the depth range of the points, the Granularity Uncertainty is introduced to Adaptive-Grained Depth Refinement stage. Granularity Uncertainty adjusts the matching range and selects the candidates' features according to coarse prediction confidence and depth range. We verify the performance of DDL-Net on SceneFlow dataset and DrivingStereo dataset by different depth metrics. Results show that DDL-Net achieves an average improvement of 25% on the SceneFlow dataset and $12\%$ on the DrivingStereo dataset comparing the classical methods. More importantly, we achieve state-of-the-art accuracy at a large distance.

SIRI: Spatial Relation Induced Network For Spatial Description Resolution

Spatial Description Resolution, as a language-guided localization task, is proposed for target location in a panoramic street view, given corresponding language descriptions. Explicitly characterizing an object-level relationship while distilling spatial relationships are currently absent but crucial to this task. Mimicking humans, who sequentially traverse spatial relationship words and objects with a first-person view to locate their target, we propose a novel spatial relationship induced (SIRI) network. Specifically, visual features are firstly correlated at an implicit object-level in a projected latent space; then they are distilled by each spatial relationship word, resulting in each differently activated feature representing each spatial relationship. Further, we introduce global position priors to fix the absence of positional information, which may result in global positional reasoning ambiguities. Both the linguistic and visual features are concatenated to finalize the target localization. Experimental results on the Touchdown show that our method is around 24\% better than the state-of-the-art method in terms of accuracy, measured by an 80-pixel radius. Our method also generalizes well on our proposed extended dataset collected using the same settings as Touchdown.

Rethink Maximum Mean Discrepancy for Domain Adaptation

Existing domain adaptation methods aim to reduce the distributional difference between the source and target domains and respect their specific discriminative information, by establishing the Maximum Mean Discrepancy (MMD) and the discriminative distances. However, they usually accumulate to consider those statistics and deal with their relationships by estimating parameters blindly. This paper theoretically proves two essential facts: 1) minimizing the MMD equals to maximize the source and target intra-class distances respectively but jointly minimize their variance with some implicit weights, so that the feature discriminability degrades; 2) the relationship between the intra-class and inter-class distances is as one falls, another rises. Based on this, we propose a novel discriminative MMD. On one hand, we consider the intra-class and inter-class distances alone to remove a redundant parameter, and the revealed weights provide their approximate optimal ranges. On the other hand, we design two different strategies to boost the feature discriminability: 1) we directly impose a trade-off parameter on the implicit intra-class distance in MMD to regulate its change; 2) we impose the similar weights revealed in MMD on inter-class distance and maximize it, then a balanced factor could be introduced to quantitatively leverage the relative importance between the feature transferability and its discriminability. The experiments on several benchmark datasets not only prove the validity of theoretical results but also demonstrate that our approach could perform better than the comparative state-of-art methods substantially.

Location-Aware Feature Selection Text Detection Network

Regression-based text detection methods have already achieved promising performances with simple network structure and high efficiency. However, they are behind in accuracy comparing with recent segmentation-based text detectors. In this work, we discover that one important reason to this case is that regression-based methods usually utilize a fixed feature selection way, i.e. selecting features in a single location or in neighbor regions, to predict components of the bounding box, such as the distances to the boundaries or the rotation angle. The features selected through this way sometimes are not the best choices for predicting every component of a text bounding box and thus degrade the accuracy performance. To address this issue, we propose a novel Location-Aware feature Selection text detection Network (LASNet). LASNet selects suitable features from different locations to separately predict the five components of a bounding box and gets the final bounding box through the combination of these components. Specifically, instead of using the classification score map to select one feature for predicting the whole bounding box as most of the existing methods did, the proposed LASNet first learn five new confidence score maps to indicate the prediction accuracy of the bounding box components, respectively. Then, a Location-Aware Feature Selection mechanism (LAFS) is designed to weightily fuse the top-$K$ prediction results for each component according to their confidence score, and to combine the all five fused components into a final bounding box. As a result, LASNet predicts the more accurate bounding boxes by using a learnable feature selection way. The experimental results demonstrate that our LASNet achieves state-of-the-art performance with single-model and single-scale testing, outperforming all existing regression-based detectors.

Sparsely-Labeled Source Assisted Domain Adaptation

Domain Adaptation (DA) aims to generalize the classifier learned from the source domain to the target domain. Existing DA methods usually assume that rich labels could be available in the source domain. However, there are usually a large number of unlabeled data but only a few labeled data in the source domain, and how to transfer knowledge from this sparsely-labeled source domain to the target domain is still a challenge, which greatly limits their application in the wild. This paper proposes a novel Sparsely-Labeled Source Assisted Domain Adaptation (SLSA-DA) algorithm to address the challenge with limited labeled source domain samples. Specifically, due to the label scarcity problem, the projected clustering is conducted on both the source and target domains, so that the discriminative structures of data could be leveraged elegantly. Then the label propagation is adopted to propagate the labels from those limited labeled source samples to the whole unlabeled data progressively, so that the cluster labels are revealed correctly. Finally, we jointly align the marginal and conditional distributions to mitigate the cross-domain mismatch problem, and optimize those three procedures iteratively. However, it is nontrivial to incorporate those three procedures into a unified optimization framework seamlessly since some variables to be optimized are implicitly involved in their formulas, thus they could not promote to each other. Remarkably, we prove that the projected clustering and conditional distribution alignment could be reformulated as different expressions, thus the implicit variables are revealed in different optimization steps. As such, the variables related to those three quantities could be optimized in a unified optimization framework and facilitate to each other, to improve the recognition performance obviously.