Custom dataset

[woodcore-an]

If I want to train my datasets(texture-less), which has 4 classes.

First, how should I train the 2D detector? And MaskRCNN or Retinanet?

Second, in surfemb, What parameters should I modify?

Can you give me some advice? Thanks~

[rasmushaugaard]

You can use any detector. MaskRCNN, RetinaNet, Yolo, etc. You don't have to change any parameters.

[rasmushaugaard]

Train a detector on your dataset, then run detection inference on your test set and save the inference results (bboxes, obj ids, etc.) in data/detection_results/[dataset].

[woodcore-an]

oh~, i get it. If I want to input an image and get its pose, or use the realsense camera for real-time pose estimation, how do I do it?

[rasmushaugaard]

Then you need to write your own inference script. I don't have time to help you with that, but you can get some inspiration from surfemb/scripts/infer.py.

[woodcore-an]

Okay, thanks. I'll have a try.

[cats0212]

Okay, thanks. I'll have a try.

Hi,Can you run it successfully "python -m surfemb.scripts.train tless" or "python -m surfemb.scripts.train ycbv", with pbr or real imgs?

[woodcore-an]

yeah, i run python -m surfemb.scripts.train tlesssuccessfully with pbr or with pbr and real.

[cats0212]

yeah, i run python -m surfemb.scripts.train tlesssuccessfully with pbr or with pbr and real.

thank you~

[cats0212]

yeah, i run python -m surfemb.scripts.train tlesssuccessfully with pbr or with pbr and real.

Hi,How many gpu do you use for training?if I use one gpu,it is all right,but when I use 2 or 4 gpus,it will can not train.

[woodcore-an]

I only use one gpu for training, the 2080ti, my other one is a 3080 and I don't know how to use them together for training. If you know, please let me know.

[cats0212]

I only use one gpu for training, the 2080ti, my other one is a 3080 and I don't know how to use them together for training. If you know, please let me know.

Sorry, my gpus is the same model

[cats0212]

I only use one gpu for training, the 2080ti, my other one is a 3080 and I don't know how to use them together for training. If you know, please let me know.

'--max-steps', type=int, default=500_000,How long did you train cnns by 2080ti?I spent 2.5 hours for 1000 steps by 3090.I feel too slow.Maybe there's something wrong with my environment.

[cats0212]

I only use one gpu for training, the 2080ti, my other one is a 3080 and I don't know how to use them together for training. If you know, please let me know.

Hi,after train in tless,do you get the bop recall( bop_toolkit/bop_toolkit_lib/eval_bop19.py) in the paper?

[woodcore-an]

I only use one gpu for training, the 2080ti, my other one is a 3080 and I don't know how to use them together for training. If you know, please let me know.

Hi,after train in tless,do you get the bop recall( bop_toolkit/bop_toolkit_lib/eval_bop19.py) in the paper?

Lower than the paper. RGB: 0.720, RGBD: 0.809.

What about you?

[cats0212]

Hi,after train in tless,do you get the bop recall( bop_toolkit/bop_toolkit_lib/eval_bop19.py) in the paper?

Lower than the paper. RGB: 0.720, RGBD: 0.809.

What about you?

I train in ycbv datasets.RGB 0.56 in Pbr.lower than the paper.I will try train in tless on the weekend, then tell you.

[cats0212]

I only use one gpu for training, the 2080ti, my other one is a 3080 and I don't know how to use them together for training. If you know, please let me know.

Hi,after train in tless,do you get the bop recall( bop_toolkit/bop_toolkit_lib/eval_bop19.py) in the paper?

Lower than the paper. RGB: 0.720, RGBD: 0.809.

What about you?

Hi,I run python -m surfemb.scripts.infer tless tless-2rs64lwh.compact.ckpt [,()](> > > I only use one gpu for training, the 2080ti, my other one is a 3080 and I don't know how to use them together for training. If you know, please let me know.

Hi,after train in tless,do you get the bop recall( bop_toolkit/bop_toolkit_lib/eval_bop19.py) in the paper?

Lower than the paper. RGB: 0.720, RGBD: 0.809.

What about you?

Hi,I run python -m surfemb.scripts.infer tless tless-2rs64lwh.compact.ckpt ,(wget https://github.com/rasmushaugaard/surfemb/releases/download/v0.0.1/tless-2rs64lwh.compact.ckpt -P data/models) Recall:0.637,Have you tried the results of using this model directly?) Recall:0.637,Have you tried the results of using this model directly?

[woodcore-an]

What do you mean by that? I do not understand very well

Hi,I run python -m surfemb.scripts.infer tless tless-2rs64lwh.compact.ckpt ,(wget https://github.com/rasmushaugaard/surfemb/releases/download/v0.0.1/tless-2rs64lwh.compact.ckpt -P data/models) Recall:0.637,Have you tried the results of using this model directly?) Recall:0.637,Have you tried the results of using this model directly?

[cats0212]

What do you mean by that? I do not understand very well

Hi,I run python -m surfemb.scripts.infer tless tless-2rs64lwh.compact.ckpt ,(wget https://github.com/rasmushaugaard/surfemb/releases/download/v0.0.1/tless-2rs64lwh.compact.ckpt -P data/models) Recall:0.637,Have you tried the results of using this model directly?) Recall:0.637,Have you tried the results of using this model directly?

The author of the paper have provided trained models([https://github.com/rasmushaugaard/surfemb/releases/download/v0.0.1/tless-2rs64lwh.compact.ckpt]). Have you tried infer with the provided model？I got recall 0.63.

[rasmushaugaard]

Just to be sure, I did a fresh install on a remote server, following the instructions in the readme, and then ran RGB inference with tless-2rs64lwh.compact.ckpt. I still get AR=0.735 like published on BOP.

[cats0212]

Just to be sure, I did a fresh install on a remote server, following the instructions in the readme, and then ran RGB inference with tless-2rs64lwh.compact.ckpt. I still get AR=0.735 like published on BOP.

OK,I try again.Thank you

[woodcore-an]

I got AR=0.7177

[cats0212]

Just to be sure, I did a fresh install on a remote server, following the instructions in the readme, and then ran RGB inference with tless-2rs64lwh.compact.ckpt. I still get AR=0.735 like published on BOP.

I got AR=0.735 in personal computer. conda env create -f environment.yml But server gpus are 3090, cuda 11.3.Personal computer 1080ti, cuda 10.2. pytorch,python,etc are same version in server and personal computer.I don't understand why it's different in recall.

[cats0212]

I got AR=0.7177

Thank you,I got AR=0.735 in personal computer.

[transcend-lzy]

I got AR=0.7177

Thank you,I got AR=0.735 in personal computer.

Hello, have you encountered this problem when training tless? Thanks for your help！ TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class 'trimesh.caching.TrackedArray'>

[yangkaihang1998]

Just to be sure, I did a fresh install on a remote server, following the instructions in the readme, and then ran RGB inference with tless-2rs64lwh.compact.ckpt. I still get AR=0.735 like published on BOP.

OK,I try again.Thank you could you tell me what problem with this (instance = self.instances[i].copy() IndexError: list index out of range) when i run the python -m surfemb.scripts.infer_debug /home/lab519/Downloads/surfemb/data/models/tless-2rs64lwh.compact.ckpt --device cuda:0 thanks