t-less datasets pose ground truth about clean-pvnet

[pengsida]

You are right. I have been performing experiments on the tless dataset.
I will upload an instruction for training the tless dataset and pretrained models.
A related code can be found at https://github.com/zju3dv/clean-pvnet/blob/master/lib/datasets/tless/handle_test_data.py#L212

[pengsida]

Training data: https://github.com/zju3dv/clean-pvnet/blob/master/lib/datasets/tless/tless_to_coco.py#L104

[firstdeep]

Thank you for your reply :)

Your research is very good.
I used the 23rd object of tless and achieved 2d projections metric 0.96 in 140th time of epoch.

I wonder if it is right to learn so quickly.
The performance is so amazing.

Thank you

[pengsida]

I am not sure.
I did not upload the evaluation code for the tless.
I just began to train the tless.

[firstdeep]

Oh, really?

I conducted the train and test using the "Training on the custom object" item you gave me.

I'll be waiting for your tless instruction.
Thank you.

[pengsida]

However, the tless has many scenes, and each scene has many instances of the same object.
I do not know what data you use for training and testing.
I will complete the training in two months.

[firstdeep]

I conducted the training using only 1296 RGB images of object 23 of Microsoft KINect v2.

The test used a tless test image.

I just want to find the posture of one object.

Of course, it depends on the orientation of the study,
but is it OK?? to use the reading "Training on the Custom object" you wrote in the README to conduct the data train?

[pengsida]

A test image?
The custom dataset is used for the case that there is only an instance of the target object in the image.

[firstdeep]

I'll check the code a little bit more.

Thank you for your kind answer.

[firstdeep]

@pengsida

Training data: https://github.com/zju3dv/clean-pvnet/blob/master/lib/datasets/tless/tless_to_coco.py#L113

I wonder how to calculate "farthest.txt" in tless datasets

thank you

[pengsida]

I will upload the dataset and pretrained models tomorrow.

[pengsida]

I sampled the farthest points using

clean-pvnet/tools/handle_custom_dataset.py

Line 19 in bbaf334

def sample_fps_points(data_root):

[firstdeep]

@pengsida

Thank you very much.

I would also appreciate it if you could tell me how to train & test the tless datasets tomorrow.
I'm having a hard time at the test.

Thank you

Have a good night.

-david

[pengsida]

I have uploaded the code, pretrained models and data about tless.

[firstdeep]

@pengsida

Thank you so much for sharing the tless code.

I am going to use the kinect camera datasets. Therefore, I would like to ask about the format of the tless datasets and how it is generated.

thank you

david

[pengsida]

I process the tless data using https://github.com/zju3dv/clean-pvnet/blob/master/run.py#L174
handle_test_data processes the test data.
handle_ag_data and tless_to_coco processes the training data for detection and pose estimation.
handle_rendering_data synthesizes images for training pose estimation.