Jiang Haiyong

Jiacong Wang, Bohong Wu, Haiyong Jiang et al.

Recent advances in Vision-Language Models (VLMs) and the scarcity of high-quality multi-modal alignment data have inspired numerous researches on synthetic VLM data generation. The conventional norm in VLM data construction uses a mixture of specialists in caption and OCR, or stronger VLM APIs and expensive human annotation. In this paper, we present World to Code (W2C), a meticulously curated multi-modal data construction pipeline that organizes the final generation output into a Python code format. The pipeline leverages the VLM itself to extract cross-modal information via different prompts and filter the generated outputs again via a consistency filtering strategy. Experiments have demonstrated the high quality of W2C by improving various existing visual question answering and visual grounding benchmarks across different VLMs. Further analysis also demonstrates that the new code parsing ability of VLMs presents better cross-modal equivalence than the commonly used detail caption ability. Our code is available at https://github.com/foundation-multimodal-models/World2Code.

Weng Tingyu, Xiao Jun, Jiang Haiyong

3D recognition is the foundation of 3D deep learning in many emerging fields, such as autonomous driving and robotics.Existing 3D methods mainly focus on the recognition of a fixed set of known classes and neglect possible unknown classes during testing. These unknown classes may cause serious accidents in safety-critical applications, i.e. autonomous driving. In this work, we make a first attempt to address 3D open-set recognition (OSR) so that a classifier can recognize known classes as well as be aware of unknown classes. We analyze open-set risks in the 3D domain and point out the overconfidence and under-representation problems that make existing methods perform poorly on the 3D OSR task. To resolve above problems, we propose a novel part prototype-based OSR method named PartCom. We use part prototypes to represent a 3D shape as a part composition, since a part composition can represent the overall structure of a shape and can help distinguish different known classes and unknown ones. Then we formulate two constraints on part prototypes to ensure their effectiveness. To reduce open-set risks further, we devise a PUFS module to synthesize unknown features as representatives of unknown samples by mixing up part composite features of different classes. We conduct experiments on three kinds of 3D OSR tasks based on both CAD shape dataset and scan shape dataset. Extensive experiments show that our method is powerful in classifying known classes and unknown ones and can attain much better results than SOTA baselines on all 3D OSR tasks. The project will be released.