reproducing results about gs-ir HOT 6 OPEN

Albedo looks fine but occlusion looks dirty and ambiguous.
The occlusion threshold can be set --occlusion 0.25 when you conduct baking.py.

from gs-ir.

Same here.

The albedo and roughness have some results though I think it is still wrong if you compute the psnr on albedo which is only around 20.

I follow this training process:

python train.py
-m outputs/lego/
-s lego
--iterations 30000
--eval

python baking.py
-m outputs/lego/
--checkpoint outputs/lego/chkpnt30000.pth
--bound 1.5
--occlu_res 128
--occlusion 0.25

python train.py
-m outputs/lego/
-s lego/
--start_checkpoint outputs/lego/chkpnt30000.pth
--iterations 35000
--eval
--gamma
--indirect

from gs-ir.

Well, the occlusion is so strange (dirty), and I've checked the code.
I think that you can set the sample_rays here as 1 and check the occlusion (This may be a mistake when I release the code):

Since I am recovering ambient occlusion from spherical harmonics, this already includes hemisphere pre-convolution, so there is no need to perform importance sampling.

from gs-ir.

Well, the occlusion is so strange (dirty), and I've checked the code. I think that you can set the sample_rays here as 1 and check the occlusion (This may be a mistake when I release the code):

GS-IR/gs-ir/gs_ir/__init__.py

Line 16 in e8d979e

sample_rays: int = 256,

Since I am recovering ambient occlusion from spherical harmonics, this already includes hemisphere pre-convolution, so there is no need to perform importance sampling.

GS-IR/gs-ir/src/occlusion_kernel.cu

Lines 184 to 237 in e8d979e

	for (uint32_t i = 0u; i < num_samples; ++i) {
	float2 Xi = Hammersley(i, num_samples);
	float3 sample_dir = importanceSampleGGX(Xi, L, roughness, eps);
	// const float phi = 2.0f * M_PIf * (Xi.x + eps);
	// // const float cosTheta = 1.0 - Xi.y; // uniform
	// const float cosTheta = sqrt(1.0 - Xi.y); // cos
	// const float sinTheta = sqrt(1.0 - cosTheta * cosTheta);
	// // from spherical coordinates to cartesian coordinates - halfway vector
	// float3 H = make_float3(cosf(phi) * sinTheta, sinf(phi) * sinTheta, cosTheta);
	// H = normalize(H);
	// float3 sample_dir = normalize(tangent * H.x + bitangent * H.y + L * H.z);
	// const uint32_t offset = i * 3;
	// curr_sample_dir[offset + 0] = sample_dir.x;
	// curr_sample_dir[offset + 1] = sample_dir.y;
	// curr_sample_dir[offset + 2] = sample_dir.z;
	const float x = sample_dir.x, y = sample_dir.y, z = sample_dir.z;
	// conduct SH reconstruction
	for (uint32_t c = 0u; c < channels; ++c) {
	float accum = 0.0f;
	const float *coeffs = curr_coeffs + c * d2; // [d2]
	accum += 0.28209479177387814f * coeffs[0];
	if (sh_degree <= 1) {
	continue;
	}
	accum += -0.4886025119029199f * y * coeffs[1];
	accum += 0.4886025119029199f * z * coeffs[2];
	accum += -0.4886025119029199f * x * coeffs[3];
	if (sh_degree <= 2) {
	continue;
	}
	const float xy = x * y, yz = y * z, xz = x * z, xx = x * x, yy = y * y, zz = z * z;
	accum += 1.0925484305920792f * xy * coeffs[4];
	accum += -1.0925484305920792f * yz * coeffs[5];
	accum += 0.31539156525252005 * (2.0 * zz - xx - yy) * coeffs[6];
	accum += -1.0925484305920792f * xz * coeffs[7];
	accum += 0.5462742152960396f * (xx - yy) * coeffs[8];
	if (sh_degree <= 3) {
	continue;
	}
	accum += -0.5900435899266435f * y * (3.0f * xx - yy) * coeffs[9];
	accum += 2.890611442640554f * xy * z * coeffs[10];
	accum += -0.4570457994644658f * y * (4.0f * zz - xx - yy) * coeffs[11];
	accum += 0.3731763325901154f * z * (2.0f * zz - 3.0f * xx - 3.0f * yy) * coeffs[12];
	accum += -0.4570457994644658f * x * (4.0f * zz - xx - yy) * coeffs[13];
	accum += 1.4453057213202769f * z * (xx - yy) * coeffs[14];
	accum += -0.5900435899266435f * x * (xx - 3.0f * yy) * coeffs[15];
	curr_output[c] += clamp(accum, 0.0f, 1.0f);
	}
	}

Hi, thanks for you reply. I wonder what do you mean "occlusion", I think this output is roughness map?

Could you be more specific about how to visualize the occlusion? Thanks!

from gs-ir.

You can directly save the occlusion as an image.

from gs-ir.

Well, the occlusion is so strange (dirty), and I've checked the code. I think that you can set the sample_rays here as 1 and check the occlusion (This may be a mistake when I release the code):

GS-IR/gs-ir/gs_ir/__init__.py

Line 16 in e8d979e

sample_rays: int = 256,

Since I am recovering ambient occlusion from spherical harmonics, this already includes hemisphere pre-convolution, so there is no need to perform importance sampling.

GS-IR/gs-ir/src/occlusion_kernel.cu

Lines 184 to 237 in e8d979e

	for (uint32_t i = 0u; i < num_samples; ++i) {
	float2 Xi = Hammersley(i, num_samples);
	float3 sample_dir = importanceSampleGGX(Xi, L, roughness, eps);
	// const float phi = 2.0f * M_PIf * (Xi.x + eps);
	// // const float cosTheta = 1.0 - Xi.y; // uniform
	// const float cosTheta = sqrt(1.0 - Xi.y); // cos
	// const float sinTheta = sqrt(1.0 - cosTheta * cosTheta);
	// // from spherical coordinates to cartesian coordinates - halfway vector
	// float3 H = make_float3(cosf(phi) * sinTheta, sinf(phi) * sinTheta, cosTheta);
	// H = normalize(H);
	// float3 sample_dir = normalize(tangent * H.x + bitangent * H.y + L * H.z);
	// const uint32_t offset = i * 3;
	// curr_sample_dir[offset + 0] = sample_dir.x;
	// curr_sample_dir[offset + 1] = sample_dir.y;
	// curr_sample_dir[offset + 2] = sample_dir.z;
	const float x = sample_dir.x, y = sample_dir.y, z = sample_dir.z;
	// conduct SH reconstruction
	for (uint32_t c = 0u; c < channels; ++c) {
	float accum = 0.0f;
	const float *coeffs = curr_coeffs + c * d2; // [d2]
	accum += 0.28209479177387814f * coeffs[0];
	if (sh_degree <= 1) {
	continue;
	}
	accum += -0.4886025119029199f * y * coeffs[1];
	accum += 0.4886025119029199f * z * coeffs[2];
	accum += -0.4886025119029199f * x * coeffs[3];
	if (sh_degree <= 2) {
	continue;
	}
	const float xy = x * y, yz = y * z, xz = x * z, xx = x * x, yy = y * y, zz = z * z;
	accum += 1.0925484305920792f * xy * coeffs[4];
	accum += -1.0925484305920792f * yz * coeffs[5];
	accum += 0.31539156525252005 * (2.0 * zz - xx - yy) * coeffs[6];
	accum += -1.0925484305920792f * xz * coeffs[7];
	accum += 0.5462742152960396f * (xx - yy) * coeffs[8];
	if (sh_degree <= 3) {
	continue;
	}
	accum += -0.5900435899266435f * y * (3.0f * xx - yy) * coeffs[9];
	accum += 2.890611442640554f * xy * z * coeffs[10];
	accum += -0.4570457994644658f * y * (4.0f * zz - xx - yy) * coeffs[11];
	accum += 0.3731763325901154f * z * (2.0f * zz - 3.0f * xx - 3.0f * yy) * coeffs[12];
	accum += -0.4570457994644658f * x * (4.0f * zz - xx - yy) * coeffs[13];
	accum += 1.4453057213202769f * z * (xx - yy) * coeffs[14];
	accum += -0.5900435899266435f * x * (xx - 3.0f * yy) * coeffs[15];
	curr_output[c] += clamp(accum, 0.0f, 1.0f);
	}
	}

你好，我更改了sample_rays int = 1然后重新跑了一下，渲染出来的图片，很奇怪，请问这样子正常吗

from gs-ir.