Mismatch between paper and code in setting transmittance below threshold to zero

[dingjr7]

Dear developers,

I found in appendix C.1 in your paper that "Lastly, we stop ray marching and set the remaining contribution
to zero as soon as the transmittance of the ray drops below 10−4".

However, in src/testbed_nerf.cu:1625

	if (src_payload.alive) {
		uint32_t idx = atomicAdd(counter, 1);
		dst_payloads[idx] = src_payload;
		dst_rgba[idx] = src_rgba[i];
	} else if (src_rgba[i].w() > 0.001f) {
		uint32_t idx = atomicAdd(finalCounter, 1);
		dst_final_payloads[idx] = src_payload;
		dst_final_rgba[idx] = src_rgba[i];
	}

which is 10^-3 not 10^-4. Hoping to hear from you!

Best regards!

[Tom94]

Hi there, that code is not actually the condition for terminating the ray.

The code that you linked checks whether anything was hit at all before the ray died by thresholding the opacity (not the transmittance) with 10^-3. Essentially, it skips compositing entirely for rays that didn't interact appreciably with anything.

The termination criterion is

		if (local_rgba.w() > (1.0f - min_alpha)) {
			local_rgba /= local_rgba.w();
			break;
		}

from composite_kernel_nerf, where min_alpha is set to 1e-4 in all the paper results, although the GUI uses 1e-2 by default for slightly higher performance.

That said, the variable min_alpha is actually misnamed... it should be called min_transmittance. Will fix. :)

Cheers!