Question about mean model points

[taeyeopl]

Q1. The mean shapes differ from SPD mean shape. I wonder how you get mean shape.

Q2. May I ask why you added a shuffle during the loading mean model points?

CATRE/core/catre/datasets/data_loader.py

Line 376 in a0b42b5

if shuffle:

[shanice-l]

Hi,
Q1: The mean shape is derived from SPD. For adjusting CATRE, we rescaled the range to [-1, 1].

Q2: There is no special reason. I've done related experiments and found that the shuffle operation won't affect the results.

[taeyeopl]

Thanks for your quick replies.

Regarding Q1, do you have any reason for rescaling the range to [-1, 1]? Does this improve the performance?

[shanice-l]

Because we also refine the estimated scale, in the code we multiply the mean shape with s_{est}.

[taeyeopl]

Thanks, Is this part what you say?

CATRE/core/catre/engine/batch_test.py

Lines 85 to 90 in a0b42b5

	tfd_kps = transform_normed_pts_batch(
	batch["obj_kps"],
	r_est,
	t=None if cfg.INPUT.ZERO_CENTER_INPUT else t_est,
	scale=s_est,
	)

[shanice-l]

Yes.

[taeyeopl]

Thanks for your replies :)

[taeyeopl]

@shanice-l Can you explain how do you get mean_scale?

CATRE/ref/nocs.py

Lines 105 to 112 in a0b42b5

	mean_scale = {
	"bottle": 0.001 * np.array([87, 220, 89], dtype=np.float32),
	"bowl": 0.001 * np.array([165, 80, 165], dtype=np.float32),
	"camera": 0.001 * np.array([88, 128, 156], dtype=np.float32),
	"can": 0.001 * np.array([68, 146, 72], dtype=np.float32),
	"laptop": 0.001 * np.array([346, 200, 335], dtype=np.float32),
	"mug": 0.001 * np.array([146, 83, 114], dtype=np.float32),
	}

The values are slightly different when I check using the mean scale from SPD mean shape.
np.abs(self.mean_model[class].max(0)) / 2

mean_scale = {
"bottle": array([0.07624906, 0.21785153, 0.07224454]),
"bowl": array([0.16847943, 0.06134908, 0.16105685])
"camera":array([0.09654312, 0.11487991, 0.16583854])
"can":array([0.11170839, 0.18665881, 0.11039044])
"laptop":array([0.1408229 , 0.10234819, 0.16517451])
"mug": array([0.18962663, 0.11490806, 0.10203232]),
}

[shanice-l]

The code is borrowed from GPV-Pose