Build the idea, one step at a time一步步建立理解
Need a math refresher?需要补一点数学基础?
Open only the tool you need; return here to use it in the next step.按需打开一个小工具,再回到本课把它用起来。
Vectors & weighted sums向量与加权求和STEP步骤 1 / 5
Coverage is only a candidate test覆盖测试只是在挑候选
Project a footprint and bound it with a rectangle. A pixel inside that rectangle is a candidate, not necessarily strongly covered. Evaluate its Gaussian weight using the inverse screen covariance and displacement from the projected center. Multiply by the evaluated opacity to get alpha.先投影覆盖范围,用矩形包住。落在矩形里的像素只是候选,不一定贡献很大。用屏幕协方差的逆矩阵和像素到中心的偏移求高斯权重,再乘当前 opacity 得到 alpha。
STEP步骤 2 / 5
Track how much is still visible记录还有多少能透过去
Start with transmittance T = 1. A layer uses fraction α of the remaining visibility and passes fraction 1 − α onward. Before layer i, multiply all earlier 1 − α factors. This is why a rear layer has a smaller effective weight.开始时透射率 T = 1。每层使用剩余可见份额中的 α,并把 1 − α 留给后面。因此第 i 层前,把前面所有 1 − α 相乘。后层的有效贡献由此减小。
STEP步骤 3 / 5
Add contributions and the background加起各层与背景贡献
Multiply each layer’s RGB by its effective weight and add the remaining background. For red alpha 0.6 in front of blue alpha 0.5, the weights are 0.6 red, 0.2 blue and 0.2 background. Swap the layers and the weights become 0.3 red, 0.5 blue and 0.2 background. The same total coverage can produce a different color.RGB 分别乘有效权重,再加剩余背景。红层 alpha 0.6 在蓝层 alpha 0.5 前面时,红、蓝、背景权重是 0.6、0.2、0.2。交换后成为 0.3、0.5、0.2。总覆盖相同,颜色却不同。
STEP步骤 4 / 5
Two equivalent traversal conventions两种顺序怎样对上
The derivation is front-to-back. Our graphics pass draws back-to-front with premultiplied over: new color = αc + (1 − α) old color. For the same layer order in space these produce the same result. Do not use front-to-back draw order with this ordinary over blend and expect the same image.上面推导按前到后。我们的图形渲染通道按后到前绘制,使用预乘 over:新颜色 = αc + (1 − α)旧颜色。对于相同的空间层序,二者结果一致。不能把前到后的绘制顺序直接塞进这个普通 over 混合。
STEP步骤 5 / 5
Make it fast without changing the story怎样加速这段计算
Tile binning collects splats relevant to a small screen region; sorting orders candidates; early termination can skip work after T becomes tiny. These are common compute-rasterizer techniques. Our baseline instead uses CPU visibility/global center-depth sorting plus GPU quads and blending. Center-depth order is an approximation for overlapping extended splats.分块收集与小屏幕区域有关的点,排序决定先后;T 很小时可提前停止。这是计算型光栅器常见的加速方法。当前基线采用 CPU 可见性筛选、全局中心深度排序,加 GPU 四边形与混合。对于有体积且相互交叠的点,中心深度排序是一种近似。
PUT NUMBERS IN代入数字
With black background, red in front gives (0.6, 0, 0.2); blue in front gives (0.3, 0, 0.5). Contributions and residual background weights sum to 1. The lab lets you check both recurrence conventions numerically.黑色背景下,红在前得到 (0.6, 0, 0.2),蓝在前得到 (0.3, 0, 0.5)。各层贡献与剩余背景权重之和是 1。实验用数字核对两种递推约定。