We propose an efficient evaluation protocol for large vision-language models (VLMs). Given their broad knowledge and reasoning capabilities, multiple benchmarks are needed for comprehensive assessment, making evaluation computationally expensive. To improve efficiency, we construct a subset that yields results comparable to full benchmark evaluations. Our benchmark classification experiments reveal that no single benchmark fully covers all challenges. We then introduce a subset construction method using farthest point sampling (FPS). Our experiments show that FPS-based benchmarks maintain a strong correlation (> 0.96) with full evaluations while using only 1% of the data. Additionally, applying FPS to an existing benchmark improves correlation with overall evaluation results, suggesting its potential to reduce unintended dataset biases.

Transfer learning is widely used as a means to leverage knowledge obtained from a source dataset to solve a downstream task on a target dataset. In the realm of image classification, models trained on large datasets often exhibit high degrees of invariance in the output spaces against intra-class variations such as geometric or chromatic changes that leave the class labels invariant. Leveraging these invariant characteristics from one model to enhance the performance of another is an attractive prospect. However, a previous study has pointed out that conventional transfer learning approaches often compromise the robustness of the pre-trained model when transferred. In this paper, we propose a novel transfer learning method called TransInv, aimed at transferring the invariant properties of a teacher model to a target model. The invariance of the teacher model is expressed by a set of augmented samples produced by our proposed data augmentation module, which is jointly optimized with the target model parameters. We demonstrate that our proposed method effectively transfers the invariant properties of the teacher model to the target model, resulting in superior model performance compared to baseline methods. The code will be released upon acceptance.

Developing foundation models for materials science has attracted attention. However, there is a lack of studies on inorganic materials due to the difficulty in the comprehensive representation of geometric concepts composing crystals: local atomic environments, their connections, and the global symmetries. We present a contrastive learning of inorganic crystal structure (CLICS) for embedding the geometric concepts, which contrasts texts representing the contextual patterns of geometries with the crystal graphs. We demonstrate that the geometric concepts are integrally embedded on the CLICS feature space, through experiments of concept retrieval from crystal graphs, similar structure search, and few-shot/imbalanced crystal structure classification.

In this work, we present Fed3DGS, a scalable 3D reconstruction framework based on 3D Gaussian splatting (3DGS) with federated learning. Existing city-scale reconstruction methods typically adopt a centralized approach, which gathers all data in a central server and reconstructs scenes. The approach hampers scalability because it places a heavy load on the server and demands extensive data storage when reconstructing scenes on a scale beyond city-scale. In pursuit of a more scalable 3D reconstruction, we propose a federated learning framework with 3DGS, which is a decentralized framework and can potentially use distributed computational resources across millions of clients. We tailor a distillation-based model update scheme for 3DGS and introduce appearance modeling for handling non-IID data in the scenario of 3D reconstruction with federated learning. We simulate our method on several large-scale benchmarks, and our method demonstrates rendered image quality comparable to centralized approaches. In addition, we also simulate our method with data collected in different seasons, demonstrating that our framework can reflect changes in the scenes and our appearance modeling captures changes due to seasonal variations.

We envision a system to continuously build and maintain a map based on earth-scale neural radiance fields (NeRF) using data collected from vehicles and drones in a lifelong learning manner. However, existing large-scale modeling by NeRF has problems in terms of scalability and maintainability when modeling earth-scale environments. Therefore, to address these problems, we propose a federated learning pipeline for large-scale modeling with NeRF. We tailor the model aggregation pipeline in federated learning for NeRF, thereby allowing local updates of NeRF. In the aggregation step, the accuracy of the clients' global pose is critical. Thus, we also propose global pose alignment to align the noisy global pose of clients before the aggregation step. In experiments, we show the effectiveness of the proposed pose alignment and the federated learning pipeline on the large-scale scene dataset, Mill19.

Ensembles of deep neural networks demonstrate improved performance over single models. For enhancing the diversity of ensemble members while keeping their performance, particle-based inference methods offer a promising approach from a Bayesian perspective. However, the best way to apply these methods to neural networks is still unclear: seeking samples from the weight-space posterior suffers from inefficiency due to the over-parameterization issues, while seeking samples directly from the function-space posterior often results in serious underfitting. In this study, we propose optimizing particles in the feature space where the activation of a specific intermediate layer lies to address the above-mentioned difficulties. Our method encourages each member to capture distinct features, which is expected to improve ensemble prediction robustness. Extensive evaluation on real-world datasets shows that our model significantly outperforms the gold-standard Deep Ensembles on various metrics, including accuracy, calibration, and robustness. Code is available at https://github.com/DensoITLab/featurePI .

Hierarchical clustering is an effective and efficient approach widely used for classical image segmentation methods. However, many existing methods using neural networks generate segmentation masks directly from per-pixel features, complicating the architecture design and degrading the interpretability. In this work, we propose a simpler, more interpretable architecture, called HCFormer. HCFormer accomplishes image segmentation by bottom-up hierarchical clustering and allows us to interpret, visualize, and evaluate the intermediate results as hierarchical clustering results. HCFormer can address semantic, instance, and panoptic segmentation with the same architecture because the pixel clustering is a common approach for various image segmentation tasks. In experiments, HCFormer achieves comparable or superior segmentation accuracy compared to baseline methods on semantic segmentation (55.5 mIoU on ADE20K), instance segmentation (47.1 AP on COCO), and panoptic segmentation (55.7 PQ on COCO).

Optimization of image transformation functions for the purpose of data augmentation has been intensively studied. In particular, adversarial data augmentation strategies, which search augmentation maximizing task loss, show significant improvement in the model generalization for many tasks. However, the existing methods require careful parameter tuning to avoid excessively strong deformations that take away image features critical for acquiring generalization. In this paper, we propose a data augmentation optimization method based on the adversarial strategy called TeachAugment, which can produce informative transformed images to the model without requiring careful tuning by leveraging a teacher model. Specifically, the augmentation is searched so that augmented images are adversarial for the target model and recognizable for the teacher model. We also propose data augmentation using neural networks, which simplifies the search space design and allows for updating of the data augmentation using the gradient method. We show that TeachAugment outperforms existing methods in experiments of image classification, semantic segmentation, and unsupervised representation learning tasks.

Superpixels are a useful representation to reduce the complexity of image data. However, to combine superpixels with convolutional neural networks (CNNs) in an end-to-end fashion, one requires extra models to generate superpixels and special operations such as graph convolution. In this paper, we propose a way to implicitly integrate a superpixel scheme into CNNs, which makes it easy to use superpixels with CNNs in an end-to-end fashion. Our proposed method hierarchically groups pixels at downsampling layers and generates superpixels. Our method can be plugged into many existing architectures without a change in their feed-forward path because our method does not use superpixels in the feed-forward path but use them to recover the lost resolution instead of bilinear upsampling. As a result, our method preserves detailed information such as object boundaries in the form of superpixels even when the model contains downsampling layers. We evaluate our method on several tasks such as semantic segmentation, superpixel segmentation, and monocular depth estimation, and confirm that it speeds up modern architectures and/or improves their prediction accuracy in these tasks.

It has been rigorously demonstrated that an end-to-end (E2E) differentiable formulation of a deep neural network can turn a complex recognition problem into a unified optimization task that can be solved by some gradient descent method. Although E2E network optimization yields a powerful fitting ability, the joint optimization of layers is known to potentially bring situations where layers co-adapt one to the other in a complex way that harms generalization ability. This work aims to numerically evaluate the generalization ability of a particular non-E2E network optimization approach known as FOCA (Feature-extractor Optimization through Classifier Anonymization) that helps to avoid complex co-adaptation, with careful hyperparameter tuning. In this report, we present intriguing empirical results where the non-E2E trained models consistently outperform the corresponding E2E trained models on three image-classification datasets. We further show that E2E network fine-tuning, applied after the feature-extractor optimization by FOCA and the following classifier optimization with the fixed feature extractor, indeed gives no improvement on the test accuracy. The source code is available at https://github.com/DensoITLab/FOCA-v1.

The paper presents a pedestrian near-miss detector with temporal analysis that provides both pedestrian detection and risk-level predictions which are demonstrated on a self-collected database. Our work makes three primary contributions: (i) The framework of pedestrian near-miss detection is proposed by providing both a pedestrian detection and risk-level assignment. Specifically, we have created a Pedestrian Near-Miss (PNM) dataset that categorizes traffic near-miss incidents based on their risk levels (high-, low-, and no-risk). Unlike existing databases, our dataset also includes manually localized pedestrian labels as well as a large number of incident-related videos. (ii) Single-Shot MultiBox Detector with Motion Representation (SSD-MR) is implemented to effectively extract motion-based features in a detected pedestrian. (iii) Using the self-collected PNM dataset and SSD-MR, our proposed method achieved +19.38% (on risk-level prediction) and +13.00% (on joint pedestrian detection and risk-level prediction) higher scores than that of the baseline SSD and LSTM. Additionally, the running time of our system is over 50 fps on a graphics processing unit (GPU).

We propose a Regularization framework based on Adversarial Transformations (RAT) for semi-supervised learning. RAT is designed to enhance robustness of the output distribution of class prediction for a given data against input perturbation. RAT is an extension of Virtual Adversarial Training (VAT) in such a way that RAT adversraialy transforms data along the underlying data distribution by a rich set of data transformation functions that leave class label invariant, whereas VAT simply produces adversarial additive noises. In addition, we verified that a technique of gradually increasing of perturbation region further improves the robustness. In experiments, we show that RAT significantly improves classification performance on CIFAR-10 and SVHN compared to existing regularization methods under standard semi-supervised image classification settings.

Multiple-object tracking (MOT) and classification are core technologies for processing moving point clouds in radar or lidar applications. For accurate object classification, the one-to-one association relationship between the model of each objects’ motion (trackers) and the observation sequences including auxiliary features (e.g., radar cross section) is important. In this work, we propose an unsupervised neural MOT model for accurate semi-automatic association labeling and we tackle the challenging one-to-one constrained combinatorial association problem by applying relaxation techniques. Experimental results demonstrate that our neural MOT model generates a more constraint-consistent association solution than conventional row-wise softmax methods.

We propose an image reconstruction method from event data, assuming the target images belong to a prespecified class like QR codes. Instead of solving the reconstruction problem in the image space, we introduce a code space that covers all the noiseless target class images and solves the reconstruction problem on it. This restriction enormously reduces the number of optimizing parameters and makes the reconstruction problem well posed and robust to noise. We demonstrate fast and robust QR-code scanning in difficult, high-speed scenes with industrial high-speed cameras and other reconstruction methods.

PointNet, which is the widely used point-wise embedding method and known as a universal approximator for continuous set functions, can process one million points per second. Nevertheless, real-time inference for the recent development of high-performing sensors is still challenging with existing neural network-based methods, including PointNet. In ordinary cases, the embedding function of PointNet behaves like a soft-indicator function that is activated when the input points exist in a certain local region of the input space. Leveraging this property, we reduce the computational costs of point-wise embedding by replacing the embedding function of PointNet with the soft-indicator function by Gaussian kernels. Moreover, we show that the Gaussian kernels also satisfy the universal approximation theorem that PointNet satisfies. In experiments, we verify that our model using the Gaussian kernels achieves comparable results to baseline methods, but with much fewer floating-point operations per sample up to 92% reduction from PointNet.

Aiming at drastic speedup for point-feature embeddings at test time, we propose a new framework that uses a pair of multi-layer perceptrons (MLP) and a lookup table (LUT) to transform point-coordinate inputs into high-dimensional features. When compared with PointNet's feature embedding part realized by MLP that requires millions of dot products, the proposed framework at test time requires no such layers of matrix-vector products but requires only looking up the nearest entities from the tabulated MLP followed by interpolation, defined over discrete inputs on a 3D lattice that is substantially arranged irregularly. We call this framework LUTI-MLP: LUT Interpolation ML that provides a way to train end-to-end irregularly tabulated MLP coupled to a LUT in a specific manner without the need for any approximation at test time. LUTI-MLP also provides significant speedup for Jacobian computation of the embedding function wrt global pose coordinate on Lie algebra $\mathfrak{se}(3)$ at test time, which could be used for point-set registration problems. After extensive evaluation using the ModelNet40, we confirmed that the LUTI-MLP even with a small (e.g., $4^3$) lattice yields performance comparable to that of the MLP while achieving significant speedup: $100\times$ for the embedding, $12\times$ for the approximate Jacobian, and $860\times$ for the canonical Jacobian.

We propose an unsupervised superpixel segmentation method by optimizing a randomly-initialized convolutional neural network (CNN) in inference time. Our method generates superpixels via CNN from a single image without any labels by minimizing a proposed objective function for superpixel segmentation in inference time. There are three advantages to our method compared with many of existing methods: (i) leverages an image prior of CNN for superpixel segmentation, (ii) adaptively changes the number of superpixels according to the given images, and (iii) controls the property of superpixels by adding an auxiliary cost to the objective function. We verify the advantages of our method quantitatively and qualitatively on BSDS500 dataset.

A method that can robustly estimate the distance to a foregoing vehicle with a monocular camera is in high demand for developing inexpensive Adaptive Cruise Control (ACC) systems. The accuracy of the estimated distances heavily depends on the road geometry and the true distance to the target vehicle. That motivates us to develop a distance estimation method that has two important properties: (a) it can deal with non-planar road surfaces from the first principle, and (b) it can estimate not only the distance but its variance of the error through mathematical analyses. Experiments on the KITTI dataset demonstrated that the proposed distance estimator is more robust and stable than the planar assumption model, and we verified by Monte Carlo simulations that our error model correctly predicts the variance of the error in the estimated distances.

The paper presents a novel concept that analyzes and visualizes worldwide fashion styles. Our goal is to reveal web-based viral fashion styles. To achieve the fashion-based analysis, we have collected fashion culture database (FCDB), which consists of 76 million geo-tagged images in 16 cosmopolitan cities. The database allows us to grasp a trend of mixed fashion styles with a fashion-based descriptor and codeword vector. In order to unveil web-based fashion trends in the FCDB, we applied a simple technique that is a temporal subtraction between consecutive codeword vectors in two different times. In the experiments, we show the analysis of fashion trends and fashion-based city similarity in a social media. As the result of large-scale data collection, we achieved world-level fashion visualization.