NAAMII NeRF Paper Reading Session

volumetric rendering

Density \(\sigma(r(t))\) is the differential probability that a ray terminates at \(t\) . \(C(r(t))\) is the color at \(r(t)\) .

Transmittance \(T(t)\) is the probability that the ray travels upto \(t\) without terminating.

\(T(t)\sigma(t)\) is the probability that the ray travels upto \(t\) , and then terminates at \(t\) . If a ray terminates at \(t\) , it receives the color \(C(t)\) .

Color of a ray \(r\) , \(C(r)\) , is the expectation of \(C(t)\) under \(T(t) \sigma(t)\) . \[C(r) = \int_{t_n}^{t_f}{T(t) \sigma(t) c(t)} dt\]

Transmittance and density are related:

\[T(t+dt) = T(t) (1 - \sigma(t)dt)\] \[T(t) + T'(t)dt = T(t) - T(t)\sigma(t)dt\] \[\frac{T'(t)}{T(t)} = - \sigma(t)dt\] \[\sigma(t) = - \frac{d}{dt}log(T(t))\] Density is the score function of transmittance! \[\log(T(t)) = - \int_{t_n}^{t}{\sigma(s)ds}\] \[T(t) = \exp \left(- \int_{t_n}^{t}{\sigma(s)ds}\right)\]

Our goal is to evaluate this double integral: \[C(r) = \int_{t_n}^{t_f}{T(t) \sigma(t) c(t)} dt\]

We need to evaluate this double integral. Let's approximate \(T(t)\) first.

\[T(t) = \exp \left(- \int_{t_n}^{t}{\sigma(s)ds}\right)\] \[= \exp \left(- \int_{t_n}^{t_i}{\sigma(s)ds}) - \int_{t_i}^{t}{\sigma(s)ds}\right)\] \[= \exp \left(- \int_{t_n}^{t_i}{\sigma(s)ds})\right) . \exp \left( - \int_{t_i}^{t}{\sigma(s)ds}\right)\] \[\approx \exp \left(- \sum_{j=1}^{i-1}{\sigma_j\delta_j}\right) . \exp \left( - \int_{t_i}^{t}{\sigma_i ds}\right)\] \[= T_i \exp \left(\sigma_i (t_i - t) \right)\]

Now, we approximate the ray color.

\[C \approx \sum_{i=1}^{n} \int_{t_i}^{t_{i+1}}{T(t)\sigma_{i}c_{i}dt}\] \[= \sum_{i=1}^{n} \int_{t_i}^{t_{i+1}}{ T_i \exp \left( \sigma_i (t_i - t) \right) \sigma_{i}c_{i}dt}\] \[= \sum_{i=1}^{n} T_i c_i \sigma_i \int_{t_i}^{t_{i+1}}{ \exp \left( \sigma_i (t_i - t) \right)dt}\] \[C \approx \hat{C} = \sum_{i=1}^{n} T_i c_i (1 - \exp(-\sigma_i \delta_i))\]