NEC Laboratories America

Tag Archive for: yi wen chen

You are here: / / yi wen chen

Yi-Wen Chen NEC Labs AmericaYi-Wen Chen is a Postdoctoral Scientist in the Media Analytics Department at NEC Laboratories America in San Jose, CA. She received her PhD in Computer Science from the University of California, Merced, where she studied multimodal machine learning and representation fusion for cross-sensor applications. She received her MS in Communication Engineering and her BS in Electrical Engineering from National Taiwan University.

At NEC, Dr. Chen’s research centers on vision–language models (VLMs), multimodal learning, and video understanding, with an emphasis on robustness, generalization, and interpretability for real-world deployments. Her recent work spans open-vocabulary / zero-shot recognition, entity grounding with LLM assistance, and text-conditioned image editing with localized control—pushing models to align fine-grained language cues with visual regions and to hold up under domain shift. She also studies temporal grounding in video, cross-modal attention/fusion for cross-sensor applications, and evaluation protocols that probe compositional reasoning and failure modes (e.g., caption drift, representation collapse, and shortcut learning). Collectively, these threads aim to make VLMs safer, more reliable, and more explainable, enabling human-centered AI for security, accessibility, and assistive scenarios.

Posts

CAViAR: A Causal Video Dataset for Fine-Grained Accident Reasoning in Real-World Scenarios

/in /by

While modern autonomous driving systems excel at perceptiontasks such as object detection and trajectory prediction, they lackthe high-level causal reasoning required to interpret traffic accidents.In particular, determining responsibility, such as identifying who is atfault and which traffic rule was violated, remains largely unexploredin current benchmarks. To this end, we introduce CAViAR (CausalAccident Video and Incident Analysis Repository), a human-annotateddashcam benchmark comprising 2,249 real-world accident videos collectedfrom CarCrashDataset (CCD) and Nexar. Each video is annotatedwith structured labels spanning environmental conditions, accidenttype, causal explanation, apparent At-Fault Agent, affected agent, andapparent rule-violation category. We benchmark state-of-the-art visionlanguagemodels (VLMs), including Cosmos-Reason2, Qwen3-VL, andInternVL3. Once class imbalance is accounted for with majority/randombaselines and balanced metrics, perceptual competence is uneven–lighting

Read more

CounterDrive: Data-Centric Adaptation of Open VLMs for Counterfactual Driving Reasoning

/in /by

Vision-language models (VLMs) have demonstrated strong capabilities in visual understanding and multimodal reasoning, yet adapting them to autonomous driving remains challenging due to the dynamic, safety-critical nature of driving environments. In particular, existing open-source VLMs often struggle to reason about hypothetical events and their potential consequences, limiting their usefulness for driving-related decision support. We present CounterDrive, a three-stage adaptation framework that enhances driving-domain counterfactual and spatiotemporal reasoning in open-source VLMs. First, we perform domain pre-alignment using mobility-oriented supervision to adapt pretrained models to driving scenes and motion dynamics. Second, we conduct mobility QA finetuning with a mixture of factual driving questions and counterfactual question-answer pairs, enabling the model to connect scene understanding with reasoning about alternative outcomes. Third, we apply preference optimization to encourage grounded, coherent, and safety-oriented responses in challenging driving scenarios. The framework is scalable and leverages automatically constructed supervision, substantially reducing the need for manual annotation. Experiments on MMERealWorld, DriveBench, and a dedicated counterfactual driving benchmark demonstrate consistent improvements over InternVL3 baselines at both 1B and 8B scales, while maintaining competitive performance on general-domain multimodal benchmarks. These results show that targeted adaptation can effectively transfer driving-domain counterfactual reasoning capabilities to compact open-source VLMs.

Read more

Celebrating the Women of NEC Laboratories America

/in /by

NEC Laboratories America celebrates Women’s History Month and International Women’s Day by recognizing the women researchers, engineers, and interns whose work in artificial intelligence, optical networking, cybersecurity, and data science is helping shape the future of technology.

Read more

Unseen Object Segmentation in Videos via Transferable Representations

/in /by

In order to learn object segmentation models in videos, conventional methods require a large amount of pixel-wise ground truth annotations. However, collecting such supervised data is time-consuming and labor-intensive. In this paper, we exploit existing annotations in source images and transfer such visual information to segment videos with unseen object categories. Without using any annotations in the target video, we propose a method to jointly mine useful segments and learn feature representations that better adapt to the target frames. The entire process is decomposed into two tasks: (1) solving a submodular function for selecting object-like segments, and (2) learning a CNN model with a transferable module for adapting seen categories in the source domain to the unseen target video. We present an iterative update scheme between two tasks to self-learn the final solution for object segmentation. Experimental results on numerous benchmark datasets show that the proposed method performs favorably against the state-of-the-art algorithms.

Read more