The explanation of dpviewer

[HibanaSledge]

Really appreciated for your contribution.

Right now I am trying to using the model to test for my own data, and with the parameter of "evs","viz","viz_flow" both set as string, there is a wrong in the terminal saying "No module named dpviewer" in devo.py.

However I cannot find any information about it. Can you please provide me something about it?

Really appreciated!

[motzet]

Hi,

The parameter viz should be set to false because we are not using the dpviewer (Pangolin). The parameter evs must be true. I will update it accordingly. You can see how to call DEVO here:

DEVO/utils/eval_utils.py

Lines 109 to 139 in 93801a4

	@torch.no_grad()
	def run_voxel(voxeldir, cfg, network, viz=False, iterator=None, timing=False, H=480, W=640, viz_flow=False, scale=1.0, **kwargs):
	slam = DEVO(cfg, network, evs=True, ht=H, wd=W, viz=viz, viz_flow=viz_flow, **kwargs)
	for i, (voxel, intrinsics, t) in enumerate(iterator):
	if timing and i == 0:
	t0 = torch.cuda.Event(enable_timing=True)
	t1 = torch.cuda.Event(enable_timing=True)
	t0.record()
	if viz:
	# import matplotlib.pyplot as plt
	# plt.switch_backend('Qt5Agg')
	visualize_voxel(voxel.detach().cpu())
	with Timer("DEVO", enabled=timing):
	slam(t, voxel, intrinsics, scale=scale)
	for _ in range(12):
	slam.update()
	poses, tstamps = slam.terminate()
	if timing:
	t1.record()
	torch.cuda.synchronize()
	dt = t0.elapsed_time(t1)/1e3
	print(f"{voxeldir}\nDEVO Network {i+1} frames in {dt} sec, e.g. {(i+1)/dt} FPS")
	flowdata = slam.flow_data if viz_flow else None
	return poses, tstamps, flowdata

When viz_flow=True, the selected patches k of the current voxel grid i are visualized, and their projected (predicted) flow onto the adjacent grids j is shown. This visualization is time-consuming.

[motzet]

I have attached a figure showing flow_viz (predicted flow visualization) of rpg_bin. Center is the current voxel grid i. The black 3x3 dots visualize the patches.

[HibanaSledge]

Ok,I see it. Really appreciated for your reply!