Scholarlink

Object pose estimation aims to obtain the pose of objects from images. In high-speed scenarios such as factory automation (FA), the latency in vision systems can cause the estimated pose to become outdated, which means the result corresponds to an earlier frame and does not match the current time, leading to control failures. This paper proposes a CNN-Transformer based 3DoF object pose prediction network that anticipates the future pose of the target object, compensating for the inevitable system delay. Temporal reasoning is enhanced via a Frame Difference Guidance module for inter-frame variation emphasis and a Dual-Multi Temporal Encoder for capturing both short-term and long-term dependencies across sampled frames. Moreover, due to the lack of suitable benchmarks, this paper contributes a video dataset, specifically designed for training and evaluating pose prediction models. The current well-trained network achieves an average pose prediction error of within 6 \(^\circ \) C under different motion scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

3DoF Object Pose Prediction with Temporal Guidance Using a CNN-Transformer Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Weicheng Lu,
  • Yuan Li,
  • Ryuji Fuchikami,
  • Takeshi Ikenaga

摘要

Object pose estimation aims to obtain the pose of objects from images. In high-speed scenarios such as factory automation (FA), the latency in vision systems can cause the estimated pose to become outdated, which means the result corresponds to an earlier frame and does not match the current time, leading to control failures. This paper proposes a CNN-Transformer based 3DoF object pose prediction network that anticipates the future pose of the target object, compensating for the inevitable system delay. Temporal reasoning is enhanced via a Frame Difference Guidance module for inter-frame variation emphasis and a Dual-Multi Temporal Encoder for capturing both short-term and long-term dependencies across sampled frames. Moreover, due to the lack of suitable benchmarks, this paper contributes a video dataset, specifically designed for training and evaluating pose prediction models. The current well-trained network achieves an average pose prediction error of within 6 \(^\circ \) C under different motion scenarios.