Junji Otsuka

Masakazu Yoshimura, Junji Otsuka, Atsushi Irie et al.

Image Signal Processors (ISPs) play important roles in image recognition tasks as well as in the perceptual quality of captured images. In most cases, experts make a lot of effort to manually tune many parameters of ISPs, but the parameters are sub-optimal. In the literature, two types of techniques have been actively studied: a machine learning-based parameter tuning technique and a DNN-based ISP technique. The former is lightweight but lacks expressive power. The latter has expressive power, but the computational cost is too heavy on edge devices. To solve these problems, we propose "DynamicISP," which consists of multiple classical ISP functions and dynamically controls the parameters of each frame according to the recognition result of the previous frame. We show our method successfully controls the parameters of multiple ISP functions and achieves state-of-the-art accuracy with low computational cost in single and multi-category object detection tasks.

Masakazu Yoshimura, Junji Otsuka, Atsushi Irie et al.

Image recognition models that work in challenging environments (e.g., extremely dark, blurry, or high dynamic range conditions) must be useful. However, creating training datasets for such environments is expensive and hard due to the difficulties of data collection and annotation. It is desirable if we could get a robust model without the need for hard-to-obtain datasets. One simple approach is to apply data augmentation such as color jitter and blur to standard RGB (sRGB) images in simple scenes. Unfortunately, this approach struggles to yield realistic images in terms of pixel intensity and noise distribution due to not considering the non-linearity of Image Signal Processors (ISPs) and noise characteristics of image sensors. Instead, we propose a noise-accounted RAW image augmentation method. In essence, color jitter and blur augmentation are applied to a RAW image before applying non-linear ISP, resulting in realistic intensity. Furthermore, we introduce a noise amount alignment method that calibrates the domain gap in the noise property caused by the augmentation. We show that our proposed noise-accounted RAW augmentation method doubles the image recognition accuracy in challenging environments only with simple training data.

Junji Otsuka, Masakazu Yoshimura, Takeshi Ohashi

Unprocessed sensor outputs (RAW images) potentially improve both low-level and high-level computer vision algorithms, but the lack of large-scale RAW image datasets is a barrier to research. Thus, reversed Image Signal Processing (ISP) which converts existing RGB images into RAW images has been studied. However, most existing methods require camera-specific metadata or paired RGB and RAW images to model the conversion, and they are not always available. In addition, there are issues in handling diverse ISPs and recovering global illumination. To tackle these limitations, we propose a self-supervised reversed ISP method that does not require metadata and paired images. The proposed method converts a RGB image into a RAW-like image taken in the same environment with the same sensor as a reference RAW image by dynamically selecting parameters of the reversed ISP pipeline based on the reference RAW image. The parameter selection is trained via pseudo paired data created from unpaired RGB and RAW images. We show that the proposed method is able to learn various reversed ISPs with comparable accuracy to other state-of-the-art supervised methods and convert unknown RGB images from COCO and Flickr1M to target RAW-like images more accurately in terms of pixel distribution. We also demonstrate that our generated RAW images improve performance on real RAW image object detection task.

Masakazu Yoshimura, Zitang Sun, Yuiko Sakuma et al.

Neural Architecture Search (NAS) aims to automatically discover high-performing deep neural network (DNN) architectures. However, conventional algorithm-driven NAS relies on carefully hand-crafted search spaces to ensure executability, which restricts open-ended exploration. Recent coding-based agentic approaches using large language models (LLMs) reduce manual design, but current LLMs struggle to reliably generate complex, valid architectures, and their proposals are often biased toward a narrow set of patterns observed in their training data. To bridge reliable algorithmic search with powerful LLM assistance, we propose LLMasTool, a hierarchical tree-based NAS framework for stable and open-ended model evolution. Our method automatically extracts reusable modules from arbitrary source code and represents full architectures as hierarchical trees, enabling evolution through reliable tree transformations rather than code generation. At each evolution step, coarse-level planning is governed by a diversity-guided algorithm that leverages Bayesian modeling to improve exploration efficiency, while the LLM resolves the remaining degrees of freedom to ensure a meaningful evolutionary trajectory and an executable generated architecture. With this formulation, instead of fully agentic LLM approaches, our method explores diverse directions beyond the inherent biases in the LLM. Our method improves over existing NAS methods by 0.69, 1.83, and 2.68 points on CIFAR-10, CIFAR-100, and ImageNet16-120, demonstrating its effectiveness.

Yuiko Sakuma, Masakazu Yoshimura, Marcel Gröpl et al.

Current neural architecture search (NAS) methods are often limited by their predefined, restrictive search spaces. While recent large language model (LLM)-assisted NAS methods enable open-ended search spaces, they often suffer from inefficient exploration due to biased or low-quality design ideas. To address these issues, we propose to semi-automatically structure model design knowledge to guide the search process. Our approach first defines a high-level structural template of architectural attributes. An LLM then populates this template by analyzing papers, creating a rich and diverse search space that embodies this structured design knowledge. To efficiently explore this vast space, we introduce FairNAD, using a multi-type mutation that enables broad exploration through mutation with fair idea sampling, Pareto-aware mutation, LLM-driven iterative mutation, and a fine-grained feedback loop. We demonstrate the effectiveness of FairNAD in discovering high-performing architectures that yield 0.84, 2.17, and 2.35 points improvement on CIFAR-10, CIFAR-100, and ImageNet16-120, respectively, compared to current state-of-the-art methods.

Siddharth Sagar Nijhawan, Leo Hoshikawa, Atsushi Irie et al.

We propose a light-weight and highly efficient Joint Detection and Tracking pipeline for the task of Multi-Object Tracking using a fully-transformer architecture. It is a modified version of TransTrack, which overcomes the computational bottleneck associated with its design, and at the same time, achieves state-of-the-art MOTA score of 73.20%. The model design is driven by a transformer based backbone instead of CNN, which is highly scalable with the input resolution. We also propose a drop-in replacement for Feed Forward Network of transformer encoder layer, by using Butterfly Transform Operation to perform channel fusion and depth-wise convolution to learn spatial context within the feature maps, otherwise missing within the attention maps of the transformer. As a result of our modifications, we reduce the overall model size of TransTrack by 58.73% and the complexity by 78.72%. Therefore, we expect our design to provide novel perspectives for architecture optimization in future research related to multi-object tracking.

Masakazu Yoshimura, Junji Otsuka, Takeshi Ohashi

Full DNN-based image signal processors (ISPs) have been actively studied and have achieved superior image quality compared to conventional ISPs. In contrast to this trend, we propose a lightweight ISP that consists of simple conventional ISP functions but achieves high image quality by increasing expressiveness. Specifically, instead of tuning the parameters of the ISP, we propose to control them dynamically for each environment and even locally. As a result, state-of-the-art accuracy is achieved on various datasets, including other tasks like tone mapping and image enhancement, even though ours is lighter than DNN-based ISPs. Additionally, our method can process different image sensors with a single ISP through dynamic control, whereas conventional methods require training for each sensor.

Christoph Reinders, Radu Berdan, Beril Besbinar et al.

Current deep learning approaches in computer vision primarily focus on RGB data sacrificing information. In contrast, RAW images offer richer representation, which is crucial for precise recognition, particularly in challenging conditions like low-light environments. The resultant demand for comprehensive RAW image datasets contrasts with the labor-intensive process of creating specific datasets for individual sensors. To address this, we propose a novel diffusion-based method for generating RAW images guided by RGB images. Our approach integrates an RGB-guidance module for feature extraction from RGB inputs, then incorporates these features into the reverse diffusion process with RGB-guided residual blocks across various resolutions. This approach yields high-fidelity RAW images, enabling the creation of camera-specific RAW datasets. Our RGB2RAW experiments on four DSLR datasets demonstrate state-of-the-art performance. Moreover, RAW-Diffusion demonstrates exceptional data efficiency, achieving remarkable performance with as few as 25 training samples or even fewer. We extend our method to create BDD100K-RAW and Cityscapes-RAW datasets, revealing its effectiveness for object detection in RAW imagery, significantly reducing the amount of required RAW images.

Sho Inayoshi, Aji Resindra Widya, Satoshi Ozaki et al.

Model pre-training has become essential in various recognition tasks. Meanwhile, with the remarkable advancements in image generation models, pre-training methods utilizing generated images have also emerged given their ability to produce unlimited training data. However, while existing methods utilizing generated images excel in classification, they fall short in more practical tasks, such as human pose estimation. In this paper, we have experimentally demonstrated it and propose the generation of visually distinct images with identical human poses. We then propose a novel multi-positive contrastive learning, which optimally utilize the previously generated images to learn structural features of the human body. We term the entire learning pipeline as GenPoCCL. Despite using only less than 1% amount of data compared to current state-of-the-art method, GenPoCCL captures structural features of the human body more effectively, surpassing existing methods in a variety of human-centric perception tasks.

Zitang Sun, Masakazu Yoshimura, Junji Otsuka et al.

High-quality data has become a primary driver of progress under scale laws, with curated datasets often outperforming much larger unfiltered ones at lower cost. Online data curation extends this idea by dynamically selecting training samples based on the model's evolving state. While effective in classification and multimodal learning, existing online sampling strategies rarely extend to object detection because of its structural complexity and domain gaps. We introduce DetGain, an online data curation method specifically for object detection that estimates the marginal perturbation of each image to dataset-level Average Precision (AP) based on its prediction quality. By modeling global score distributions, DetGain efficiently estimates the global AP change and computes teacher-student contribution gaps to select informative samples at each iteration. The method is architecture-agnostic and minimally intrusive, enabling straightforward integration into diverse object detection architectures. Experiments on the COCO dataset with multiple representative detectors show consistent improvements in accuracy. DetGain also demonstrates strong robustness under low-quality data and can be effectively combined with knowledge distillation techniques to further enhance performance, highlighting its potential as a general and complementary strategy for data-efficient object detection.

Masakazu Yoshimura, Junji Otsuka, Radu Berdan et al.

DNN-based methods have been successful in Image Signal Processor (ISP) and image enhancement (IE) tasks. However, the cost of creating training data for these tasks is considerably higher than for other tasks, making it difficult to prepare large-scale datasets. Also, creating personalized ISP and IE with minimal training data can lead to new value streams since preferred image quality varies depending on the person and use case. While semi-supervised learning could be a potential solution in such cases, it has rarely been utilized for these tasks. In this paper, we realize semi-supervised learning for ISP and IE leveraging a RAW image reconstruction (sRGB-to-RAW) method. Although existing sRGB-to-RAW methods can generate pseudo-RAW image datasets that improve the accuracy of RAW-based high-level computer vision tasks such as object detection, their quality is not sufficient for ISP and IE tasks that require precise image quality definition. Therefore, we also propose a sRGB-to-RAW method that can improve the image quality of these tasks. The proposed semi-supervised learning with the proposed sRGB-to-RAW method successfully improves the image quality of various models on various datasets.

Yuiko Sakuma, Masakazu Yoshimura, Junji Otsuka et al.

Compression of large and performant vision foundation models (VFMs) into arbitrary bit-wise operations (BitOPs) allows their deployment on various hardware. We propose to fine-tune a VFM to a mixed-precision quantized supernet. The supernet-based neural architecture search (NAS) can be adopted for this purpose, which trains a supernet, and then subnets within arbitrary hardware budgets can be extracted. However, existing methods face difficulties in optimizing the mixed-precision search space and incurring large memory costs during training. To tackle these challenges, first, we study the effective search space design for fine-tuning a VFM by comparing different operators (such as resolution, feature size, width, depth, and bit-widths) in terms of performance and BitOPs reduction. Second, we propose memory-efficient supernet training using a low-rank adapter (LoRA) and a progressive training strategy. The proposed method is evaluated for the recently proposed VFM, Segment Anything Model, fine-tuned on segmentation tasks. The searched model yields about a 95% reduction in BitOPs without incurring performance degradation.