Sai Ma

Eduardo Pérez-Pellitero, Sibi Catley-Chandar, Richard Shaw et al.

This paper reviews the challenge on constrained high dynamic range (HDR) imaging that was part of the New Trends in Image Restoration and Enhancement (NTIRE) workshop, held in conjunction with CVPR 2022. This manuscript focuses on the competition set-up, datasets, the proposed methods and their results. The challenge aims at estimating an HDR image from multiple respective low dynamic range (LDR) observations, which might suffer from under- or over-exposed regions and different sources of noise. The challenge is composed of two tracks with an emphasis on fidelity and complexity constraints: In Track 1, participants are asked to optimize objective fidelity scores while imposing a low-complexity constraint (i.e. solutions can not exceed a given number of operations). In Track 2, participants are asked to minimize the complexity of their solutions while imposing a constraint on fidelity scores (i.e. solutions are required to obtain a higher fidelity score than the prescribed baseline). Both tracks use the same data and metrics: Fidelity is measured by means of PSNR with respect to a ground-truth HDR image (computed both directly and with a canonical tonemapping operation), while complexity metrics include the number of Multiply-Accumulate (MAC) operations and runtime (in seconds).

Sichao Li, Sai Ma, Daniel Kilov et al.

LLMs and agentic systems are increasingly deployed in social environments, making normative competence critical for safe and appropriate behavior. However, existing approaches either assess normative judgment in text alone or reduce it to choosing among a fixed set of candidate actions. We argue both are insufficient. In practice, agents are never handed a menu of options; they must identify a reasonable action from scratch, grounded in visible facts and supported by inspectable reasons. We introduce NoRA, a visual first-person video benchmark that requires models to generate candidate next actions and justify each through an explicit fact-reason-action support graph. The benchmark comprises 1,420 annotated video clips, including HumanGold-190 and LLMSilver-1230 splits. Each instance is evaluated through action alignment, factual grounding, and support binding, aggregated into a single grounded reasonableness score. We benchmark 12 multimodal systems under direct, deliberate, and structured prompting regimes, finding that current VLMs frequently recover plausible actions and relevant scene facts, but consistently struggle to construct the full reasonable action space and bind selected actions to the correct local support. NoRA makes this gap measurable, shifting the evaluation question from whether a model can pick an action to whether it can justify an appropriate action for the right visible reasons.

Sai Ma, Zhuang Li, Sichao Li et al.

Earth Observation (EO) analysis is inherently interactive: resolving uncertainty often requires expanding the region of interest, retrieving historical observations, and switching across sensors such as optical and Synthetic Aperture Radar. However, most EO benchmarks collapse this process into fixed-input, single-turn tasks. To address this gap, we present EO-Gym, a controlled executable framework for multimodal, tool-using EO agents that formulates EO analysis as a Gymnasium-style local geospatial workspace backed by more than 660k multimodal files indexed by location, time, and sensor type, with 35 EO-specialized tools spanning six task families. Built on this environment, we construct EO-Gym-Data, a benchmark of 9,078 trajectories and 34,604 reasoning steps, and grounded in eight public EO datasets together with Landsat and Sentinel-2 imagery. Evaluating $10$ open and closed VLMs shows that strong general-purpose models still struggle with interactive EO reasoning, especially on temporal and cross-modal workflows. As a reference baseline, EO-Gym-4B, obtained by fine-tuning Qwen3-VL-4B-Instruct on EO-Gym-Data, improves overall Pass@3 from $0.49$ to $0.74$ under the main evaluation setting. O-Gym provides a reproducible environment for interactive EO agents, operationalizing EO as an evidence-gathering problem that requires planning across geospatial, temporal, and sensing modality.

Sai Ma, Zhuang Li, John A Taylor

Vision language models (VLMs) that enable natural language interaction with satellite imagery can democratize Earth observation by accelerating expert workflows, making data accessible to non-specialists, and enabling planet-scale automation. However, existing datasets focus mainly on short-term, high-resolution imagery from a limited number of satellites, overlooking low-resolution, multi-satellite, long-term archives, such as Landsat, that are essential for affordable and bias-robust global monitoring. We address this gap with Landsat30-AU, a large-scale vision-language dataset built from 30-meter resolution imagery collected by four Landsat satellites (5, 7, 8, and 9) over Australia, spanning more than 36 years. The dataset includes two components: Landsat30-AU-Cap, containing $196,262$ image-caption pairs, and Landsat30-AU-VQA, comprising 17,725 human-verified visual question answering (VQA) samples across eight remote sensing domains. Both datasets are curated through a bootstrapped pipeline that leverages generic VLMs with iterative refinement and human verification to ensure quality. Our evaluation of eight VLMs on our benchmark reveals that off-the-shelf models struggle to understand satellite imagery. The open-source remote-sensing VLM EarthDial achieves only 0.07 SPIDEr in captioning and a VQA accuracy of 0.48, highlighting the limitations of current approaches. Encouragingly, lightweight fine-tuning of Qwen2.5-VL-7B on Landsat30-AU improves captioning performance from 0.11 to 0.31 SPIDEr and boosts VQA accuracy from 0.74 to 0.87. Code and data are available at https://github.com/papersubmit1/landsat30-au.

Ke Wang, Sai Ma, Junlan Chen et al.

Recently, the philosophy of visual saliency and attention has started to gain popularity in the robotics community. Therefore, this paper aims to mimic this mechanism in SLAM framework by using saliency prediction model. Comparing with traditional SLAM that treated all feature points as equal important in optimization process, we think that the salient feature points should play more important role in optimization process. Therefore, we proposed a saliency model to predict the saliency map, which can capture both scene semantic and geometric information. Then, we proposed Salient Bundle Adjustment by using the value of saliency map as the weight of the feature points in traditional Bundle Adjustment approach. Exhaustive experiments conducted with the state-of-the-art algorithm in KITTI and EuRoc datasets show that our proposed algorithm outperforms existing algorithms in both indoor and outdoor environments. Finally, we will make our saliency dataset and relevant source code open-source for enabling future research.

Zhongliang Yang, Ke Wang, Sai Ma et al.

In order to promote the rapid development of image steganalysis technology, in this paper, we construct and release a multivariable large-scale image steganalysis dataset called IStego100K. It contains 208,104 images with the same size of 1024*1024. Among them, 200,000 images (100,000 cover-stego image pairs) are divided as the training set and the remaining 8,104 as testing set. In addition, we hope that IStego100K can help researchers further explore the development of universal image steganalysis algorithms, so we try to reduce limits on the images in IStego100K. For each image in IStego100K, the quality factors is randomly set in the range of 75-95, the steganographic algorithm is randomly selected from three well-known steganographic algorithms, which are J-uniward, nsF5 and UERD, and the embedding rate is also randomly set to be a value of 0.1-0.4. In addition, considering the possible mismatch between training samples and test samples in real environment, we add a test set (DS-Test) whose source of samples are different from the training set. We hope that this test set can help to evaluate the robustness of steganalysis algorithms. We tested the performance of some latest steganalysis algorithms on IStego100K, with specific results and analysis details in the experimental part. We hope that the IStego100K dataset will further promote the development of universal image steganalysis technology. The description of IStego100K and instructions for use can be found at https://github.com/YangzlTHU/IStego100K

Junxian Li, Xinyue Xu, Sai Ma et al.

Unfaithfulness remains a persistent challenge for large language models (LLMs), which often produce plausible yet ungrounded reasoning chains that diverge from perceptual evidence or final conclusions. We distinguish between behavioral faithfulness (alignment between reasoning and output) and perceptual faithfulness (alignment between reasoning and input), and introduce FaithEval for quantifying step-level and chain-level faithfulness by evaluating whether each claimed object is visually supported by the image. Building on these insights, we propose FaithAct, a faithfulness-first planning and acting framework that enforces evidential grounding at every reasoning step. Experiments across multiple reasoning benchmarks demonstrate that FaithAct improves perceptual faithfulness by up to 26% without degrading task accuracy compared to prompt-based and tool-augmented baselines. Our analysis shows that treating faithfulness as a guiding principle not only mitigates hallucination but also leads to more stable reasoning trajectories. This work thereby establishes a unified framework for both evaluating and enforcing faithfulness in multimodal reasoning.

Ke Wang, Sai Ma, Junlan Chen et al.

Visual odometry (VO) is a prevalent way to deal with the relative localization problem, which is becoming increasingly mature and accurate, but it tends to be fragile under challenging environments. Comparing with classical geometry-based methods, deep learning-based methods can automatically learn effective and robust representations, such as depth, optical flow, feature, ego-motion, etc., from data without explicit computation. Nevertheless, there still lacks a thorough review of the recent advances of deep learning-based VO (Deep VO). Therefore, this paper aims to gain a deep insight on how deep learning can profit and optimize the VO systems. We first screen out a number of qualifications including accuracy, efficiency, scalability, dynamicity, practicability, and extensibility, and employ them as the criteria. Then, using the offered criteria as the uniform measurements, we detailedly evaluate and discuss how deep learning improves the performance of VO from the aspects of depth estimation, feature extraction and matching, pose estimation. We also summarize the complicated and emerging areas of Deep VO, such as mobile robots, medical robots, augmented reality and virtual reality, etc. Through the literature decomposition, analysis, and comparison, we finally put forward a number of open issues and raise some future research directions in this field.

Sai Ma, Qingxiao Guan, Xianfeng Zhao et al.

Explanation from Sai Ma: The experiments in this paper are conducted on Caffe framework. In Caffe, there is an API to directly set the gradient in Matlab. I wrongly use it to control the 'probability', in fact, I modify the gradient directly. The misusage of API leads to wrong experiment results, and wrong theoretical analysis. Apologize to readers who have read this paper. We have submitted a correct version of this paper to Multimedia Tools and Applications and it is under revision. Thanks to Dr. Patrick Bas, who is the Associate Editor of TIFS and the anonymous reviewers of this paper. Thanks to Tingting Song from Sun Yat-sen University. We discussed some problems of this paper. Her advice helps me to improve the submitted paper to Multimedia Tools and Applications.

Sai Ma, Qingxiao Guan, Xianfeng Zhao et al.

Recently, the application of deep learning in steganalysis has drawn many researchers' attention. Most of the proposed steganalytic deep learning models are derived from neural networks applied in computer vision. These kinds of neural networks have distinguished performance. However, all these kinds of back-propagation based neural networks may be cheated by forging input named the adversarial example. In this paper we propose a method to generate steganographic adversarial example in order to enhance the steganographic security of existing algorithms. These adversarial examples can increase the detection error of steganalytic CNN. The experiments prove the effectiveness of the proposed method.