Open yuyuyu223 opened 3 months ago
Why wasn't normalization applied to rand_rays_d when calculating the visibility loss? I noticed that normalized direction were used for spherical harmonic calculations in other attributes.
if opt.lambda_visibility > 0: num = 10000 means3D = pc.get_xyz visibility = pc.get_visibility normal = pc.get_normal opacity = pc.get_opacity rand_idx = torch.randperm(means3D.shape[0])[:num] rand_visibility_shs_view = visibility.transpose(1, 2).view(-1, 1, 4 ** 2)[rand_idx] rand_rays_o = means3D[rand_idx] rand_rays_d = torch.randn_like(rand_rays_o) cov_inv = pc.get_inverse_covariance() rand_normal = normal[rand_idx] mask = (rand_rays_d * rand_normal).sum(-1) < 0 rand_rays_d[mask] *= -1 sample_sh2vis = eval_sh(3, rand_visibility_shs_view, rand_rays_d) sample_vis = torch.clamp(sample_sh2vis + 0.5, 0.0, 1.0) raytracer = RayTracer(means3D, pc.get_scaling, pc.get_rotation) trace_results = raytracer.trace_visibility( rand_rays_o, rand_rays_d, means3D, cov_inv, opacity, normal) rand_ray_visibility = trace_results["visibility"] loss_visibility = F.l1_loss(rand_ray_visibility, sample_vis) tb_dict["loss_visibility"] = loss_visibility.item() loss = loss + opt.lambda_visibility * loss_visibility
Why wasn't normalization applied to rand_rays_d when calculating the visibility loss? I noticed that normalized direction were used for spherical harmonic calculations in other attributes.