SocialGoodAI
commented
2 years ago Hi Bingo,
The SPD results report in CRNet are trained under the same setting with CRNet, and SPD results report in SGPA are trained under the same setting with SGPA for a fair comparison. Thus the batch size, training iterations, and epochs are different, which causes the problem. You can refer to our codes and papers for detailed settings.
同学您好！最近看到了您课题组内两篇有关category-level pose estimation的同期工作， [1] Category-Level 6D Object Pose Estimation via Cascaded Relation and Recurrent Reconstruction Networks. IROS 2021. [2] SGPA: Structure-Guided Prior Adaptation for Category-Level 6D Object Pose Estimation. ICCV2021.
但是有个问题令我感觉十分困惑：这两篇论文都对baseline方法: SPD进行了重新训练, 但是两篇论文中报告的重新训练结果却不一致（如下图，CAMERA数据集：ICCV中SDP*都比IROS中报告的更高，而REAL数据集却是IROS中报告的结果比ICCV的高），请问这是因为IROS论文和ICCV论文用了不同的GPU来重新训练SPD吗？ 这样的话，CR-Net和SGPA是否也是在不同的机器上训练的呢？