large drop of performance when using icp

[LeroyChou]

Dear Authors,

I hope this message finds you well. I am writing to express my appreciation for your remarkable work. I have been utilizing your pretrained weights of LM-O in an attempt to reproduce the results presented in your paper. My results closely align with yours, with a minor difference (75.4 vs. 76.9).

While examining your scripts, I noticed the argument “--use_icp True”. I was intrigued by the potential impact of using ICP to refine the pose. However, when I incorporated “--use_icp True” during evaluation, I observed a significant drop in the results ( can, driller, glue ). I am curious to know if you have previously encountered this issue and if you could possibly shed some light on why this might be happening.

Here are the results for your reference:

Method	Ape	Can	Cat	Driller	Duck	Eggbox	Glue	HolePuncher	Mean
ZebraPose(in your paper)	57.9	95.0	60.6	94.8	64.5	70.9	88.7	83.0	76.9
ZebraPose(reproduced)	55.0	94.5	57.9	93.2	61.7	70.6	89.0	81.2	75.4
ZebraPose+ICP(reproduced)	60.4	57.4	62.1	62.2	63.8	73.8	69.9	88.4	68.2

Yours, Leroy Chou

[suyz526]

Hello,

I don't have the ADD score for the results with ICP. According to the table, it doesn't work for some objects. Maybe you could first check if the renderer works well for those objects.

Best, Yongzhi