Regard trainning my object

[Okabe-Rintarou-0]

Hello, I am training my customized object, i have generated some datasets(use nb_objects to generate many instances in one scene): after some time of training, i got some results but find it strange:

I am new to DOPE, so i have no idea what's going on, can somebody give me a hand please?

[Okabe-Rintarou-0]

here are some images generated by using --save flag

[mintar]

Did you specify the symmetries? See here: https://github.com/NVlabs/Deep_Object_Pose/tree/master/scripts/nvisii_data_gen#handling-objects-with-symmetries

Also please show a visualization of the belief maps.

[TontonTremblay]

To echo @mintar you need to specify the symmetries to get stability.

[Okabe-Rintarou-0]

@TontonTremblay @mintar Thank you for your replies! Actually I have specified the symmetries through model_info.json:

{
  "symmetries_continuous": [{"axis": [0, 1, 0], "offset": [0, 0, 0]}],
  "align_axes": [{"object": [0, 0, 1], "camera": [0, 0, 1]}, {"object": [0, 1, 0], "camera": [0, 1, 0]}]
}

Here is the obj file, the y-axis is the axis of rotational symmetry: tube.obj

Here is another example with its belief map:

[Okabe-Rintarou-0]

Is it because my model lacks training?

[Okabe-Rintarou-0]

I want to detect the tube cap. I am wondering whether my 3d model is real enough?

[Okabe-Rintarou-0]

btw, i use --scale to enlarge the object(otherwise it'd be very small)