How to reconstruct the trajectory when inference?

[inhwanbae]

How to reconstruct the trajectory when inference?

[InhwanBae]

Hi @chenjx1005,

Thank you for showing interest in my paper. In the EigenTrajectory model, the anchor is created once before the training phase, and the network selects a trajectory with a minimal difference to learn how to refine it (because the ground truth is not multimodal and only provides one real path). During testing, all 20 trajectories included in the anchor are refined and evaluated, following standard benchmarks.

I am organizing the code for release, and it will be made public soon. It will probably be easier to understand when you see the code.

[InhwanBae]

Hi @chenjx1005, I just released the code. Checking the ./utils/trainer.py will be helpful in understanding the evaluation method.

EigenTrajectory/utils/trainer.py

Lines 249 to 264 in a6c8a3b

	@torch.no_grad()
	def test(self):
	self.model.eval()
	self.reset_metric()
	for batch in tqdm(self.loader_test, desc=f"Test {self.hyper_params.dataset.upper()} scene"):
	obs_traj, pred_traj = [tensor.cuda(non_blocking=True) for tensor in batch[:2]]
	output = self.model(obs_traj)
	# Evaluate trajectories
	for metric in self.stats_func.keys():
	value = self.stats_func[metric](output["recon_traj"], pred_traj)
	self.stats_meter[metric].extend(value)
	return {x: self.stats_meter[x].mean() for x in self.stats_meter.keys()}