Zechen Li

Zechen Li, Anders Søgaard

Synthetic datasets have successfully been used to probe visual question-answering datasets for their reasoning abilities. CLEVR (johnson2017clevr), for example, tests a range of visual reasoning abilities. The questions in CLEVR focus on comparisons of shapes, colors, and sizes, numerical reasoning, and existence claims. This paper introduces a minimally biased, diagnostic visual question-answering dataset, QLEVR, that goes beyond existential and numerical quantification and focus on more complex quantifiers and their combinations, e.g., asking whether there are more than two red balls that are smaller than at least three blue balls in an image. We describe how the dataset was created and present a first evaluation of state-of-the-art visual question-answering models, showing that QLEVR presents a formidable challenge to our current models. Code and Dataset are available at https://github.com/zechenli03/QLEVR

Zechen Li, Shohreh Deldari, Linyao Chen et al.

We introduce SensorLLM, a two-stage framework that enables Large Language Models (LLMs) to perform human activity recognition (HAR) from sensor time-series data. Despite their strong reasoning and generalization capabilities, LLMs remain underutilized for motion sensor data due to the lack of semantic context in time-series, computational constraints, and challenges in processing numerical inputs. SensorLLM addresses these limitations through a Sensor-Language Alignment stage, where the model aligns sensor inputs with trend descriptions. Special tokens are introduced to mark channel boundaries. This alignment enables LLMs to capture numerical variations, channel-specific features, and data of varying durations, without requiring human annotations. In the subsequent Task-Aware Tuning stage, we refine the model for HAR classification, achieving performance that matches or surpasses state-of-the-art methods. Our results demonstrate that SensorLLM evolves into an effective sensor learner, reasoner, and classifier through human-intuitive Sensor-Language Alignment, generalizing across diverse HAR datasets. We believe this work establishes a foundation for future research on time-series and text alignment, paving the way for foundation models in sensor data analysis. Our codes are available at https://github.com/zechenli03/SensorLLM.

Baiyu Chen, Wilson Wongso, Zechen Li et al.

Egocentric video-based models capture rich semantic information and have demonstrated strong performance in human activity recognition (HAR). However, their high power consumption, privacy concerns, and dependence on lighting conditions limit their feasibility for continuous on-device recognition. In contrast, inertial measurement unit (IMU) sensors offer an energy-efficient and privacy-preserving alternative, yet they suffer from limited large-scale annotated datasets, leading to weaker generalization in downstream tasks. To bridge this gap, we propose COMODO, a cross-modal self-supervised distillation framework that transfers rich semantic knowledge from the video modality to the IMU modality without requiring labeled annotations. COMODO leverages a pretrained and frozen video encoder to construct a dynamic instance queue, aligning the feature distributions of video and IMU embeddings. By distilling knowledge from video representations, our approach enables the IMU encoder to inherit rich semantic information from video while preserving its efficiency for real-world applications. Experiments on multiple egocentric HAR datasets demonstrate that COMODO consistently improves downstream classification performance, achieving results comparable to or exceeding fully supervised fine-tuned models. Moreover, COMODO exhibits strong cross-dataset generalization. Benefiting from its simplicity, our method is also generally applicable to various video and time-series pre-trained models, offering the potential to leverage more powerful teacher and student foundation models in future research. The code is available at https://github.com/Breezelled/COMODO .

Xin Hu, Haomiao Ni, Yunbei Zhang et al.

Vision language models (VLMs) have achieved remarkable success in broad visual understanding, yet they remain challenged by object-centric reasoning on rare objects due to the scarcity of such instances in pretraining data. While prior efforts alleviate this issue by retrieving additional data or introducing stronger vision encoders, these methods are still computationally intensive during finetuning VLMs and don't fully exploit the original training data. In this paper, we introduce an efficient plug-and-play module that substantially improves VLMs' reasoning over rare objects by refining visual tokens and enriching input text prompts, without VLMs finetuning. Specifically, we propose to learn multi-modal class embeddings for rare objects by leveraging prior knowledge from vision foundation models and synonym-augmented text descriptions, compensating for limited training examples. These embeddings refine the visual tokens in VLMs through a lightweight attention-based enhancement module that improves fine-grained object details. In addition, we use the learned embeddings as object-aware detectors to generate informative hints, which are injected into the text prompts to help guide the VLM's attention toward relevant image regions. Experiments on two benchmarks show consistent and substantial gains for pretrained VLMs in rare object recognition and reasoning. Further analysis reveals how our method strengthens the VLM's ability to focus on and reason about rare objects.

Zechen Li, Weiming Huang, Kai Zhao et al.

Recently, learning urban region representations utilizing multi-modal data (information views) has become increasingly popular, for deep understanding of the distributions of various socioeconomic features in cities. However, previous methods usually blend multi-view information in a posteriors stage, falling short in learning coherent and consistent representations across different views. In this paper, we form a new pipeline to learn consistent representations across varying views, and propose the multi-view Contrastive Prediction model for urban Region embedding (ReCP), which leverages the multiple information views from point-of-interest (POI) and human mobility data. Specifically, ReCP comprises two major modules, namely an intra-view learning module utilizing contrastive learning and feature reconstruction to capture the unique information from each single view, and inter-view learning module that perceives the consistency between the two views using a contrastive prediction learning scheme. We conduct thorough experiments on two downstream tasks to assess the proposed model, i.e., land use clustering and region popularity prediction. The experimental results demonstrate that our model outperforms state-of-the-art baseline methods significantly in urban region representation learning.

Du Yin, Jinliang Deng, Shuang Ao et al.

Training models on spatio-temporal (ST) data poses an open problem due to the complicated and diverse nature of the data itself, and it is challenging to ensure the model's performance directly trained on the original ST data. While limiting the variety of training data can make training easier, it can also lead to a lack of knowledge and information for the model, resulting in a decrease in performance. To address this challenge, we presented an innovative paradigm that incorporates three separate forms of curriculum learning specifically targeting from spatial, temporal, and quantile perspectives. Furthermore, our framework incorporates a stacking fusion module to combine diverse information from three types of curriculum learning, resulting in a strong and thorough learning process. We demonstrated the effectiveness of this framework with extensive empirical evaluations, highlighting its better performance in addressing complex ST challenges. We provided thorough ablation studies to investigate the effectiveness of our curriculum and to explain how it contributes to the improvement of learning efficiency on ST data.