updated to median depth

README.md

@@ -5,13 +5,16 @@ However, it has been edited by Jonathon Luiten to also render 'depth' as well as
 
 This is needed for Jonathon's Dynamic 3D Gaussians work which can be found here: http://dynamic3dgaussians.github.io
 
-By default, the depth is calculated as 'mean depth', but there is commented out code which also calculates 'median depth' if preferred.
+By default, the depth is calculated as 'median depth', where the depth is the depth of the Gaussian center which causes the accumulated rays transmittance to drop below 0.5.
+If a ray doesn't reach this threshold it is given a default depth of 15. This median depth avoids the depth floaters around depth boundaries that 'mean depth' would give.
+If 'mean depth' is preffered, there is commented out code which also calculates 'mean depth'.
+See lines 307-308 and 363-372 of cuda_rasterizer/forward.cu.
 
 Note that the backward pass for the depth has not been implemented, so it won't work for training with depth ground-truth.
 
 Note that the code in this repo follows the (non commercial) license of Inria as laid out in LICENSE.md
 
-If you're using this as part of the Dynamic Gaussian code, just follow the installation instruction for that codebase.
+If you're using this as part of the Dynamic 3D Gaussians code, just follow the installation instruction for that codebase.
 
 To install this stand-alone I have been doing the following (although I don't think this is necessarily the best way):
 ```

cuda_rasterizer/forward.cu

@@ -304,8 +304,8 @@ renderCUDA(
 	uint32_t contributor = 0;
 	uint32_t last_contributor = 0;
 	float C[CHANNELS] = { 0 };
-	float D = 0.0f;  // Mean Depth
-//     float D = 1000.0f;  // Median Depth. TODO: This is a hack setting max_depth to 1000
+// 	float D = 0.0f;  // Mean Depth
+    float D = 15.0f;  // Median Depth. TODO: This is a hack setting max_depth to 15
 
 	// Iterate over batches until all done or range is complete
 	for (int i = 0; i < rounds; i++, toDo -= BLOCK_SIZE)
@@ -361,15 +361,15 @@ renderCUDA(
 				C[ch] += features[collected_id[j] * CHANNELS + ch] * alpha * T;
 
             // Mean depth:
-            float dep = collected_depth[j];
-            D += dep * alpha * T;
-
-//             // Median depth:
-//             if (T > 0.5f && test_T < 0.5)
-// 			{
-// 			    float dep = collected_depth[j];
-// 				D = dep;
-// 			}
+//             float dep = collected_depth[j];
+//             D += dep * alpha * T;
+
+            // Median depth:
+            if (T > 0.5f && test_T < 0.5)
+			{
+			    float dep = collected_depth[j];
+				D = dep;
+			}
 
 
 			T = test_T;