From b0f08fb1caea88c63968eb2a91b44c723470d251 Mon Sep 17 00:00:00 2001
From: Nicolas Pope <nicolas.pope@utu.fi>
Date: Mon, 12 Aug 2019 08:04:21 +0300
Subject: [PATCH] Implements #151 alternate minimums
---
applications/reconstruct/src/dibr.cu | 48 ++++++++++++++++++++++++----
1 file changed, 41 insertions(+), 7 deletions(-)
diff --git a/applications/reconstruct/src/dibr.cu b/applications/reconstruct/src/dibr.cu
index 141c5dc81..8d21279f1 100644
--- a/applications/reconstruct/src/dibr.cu
+++ b/applications/reconstruct/src/dibr.cu
@@ -399,6 +399,8 @@ __device__ inline float warpMin(float e) {
return e;
}
+#define ENERGY_THRESHOLD 0.1f
+
/*
* Pass 2: Determine depth buffer with enough accuracy for a visibility test in pass 2.
@@ -418,6 +420,12 @@ __device__ inline float warpMin(float e) {
const int x = (blockIdx.x*blockDim.x + threadIdx.x) / WARP_SIZE;
const int y = blockIdx.y*blockDim.y + threadIdx.y;
+ // Starting point for surface minimum
+ float clusterBase = params.camera.m_sensorDepthWorldMin;
+
+ // Loop to a deeper surface if not on the first one selected...
+ while (clusterBase < params.camera.m_sensorDepthWorldMax) {
+
const int lane = tid % WARP_SIZE;
if (lane == 0) {
minimum[warp] = 100000000;
@@ -428,6 +436,8 @@ __device__ inline float warpMin(float e) {
__syncwarp();
// Search for a valid minimum neighbour
+ // TODO: Should this really be minimum or the median of a depth cluster?
+ // cluster median seems very hard to calculate...
for (int i=lane; i<NEIGHBOR_WINDOW; i+=WARP_SIZE) {
const int u = (i % (2*NEIGHBOR_RADIUS_2+1)) - NEIGHBOR_RADIUS_2;
const int v = (i / (2*NEIGHBOR_RADIUS_2+1)) - NEIGHBOR_RADIUS_2;
@@ -435,7 +445,11 @@ __device__ inline float warpMin(float e) {
const float3 camPos = params.camera.kinectDepthToSkeleton(x, y, point.z);
// If it is close enough...
- if (point.z > params.camera.m_sensorDepthWorldMin && point.z < params.camera.m_sensorDepthWorldMax && length(point - camPos) <= 0.02f) {
+ // TODO: smoothing / strength should be determined by a number of factors including:
+ // 1) Depth from original source
+ // 2) Colour contrast in underlying RGB
+ // 3) Estimated noise levels in depth values
+ if (point.z > clusterBase && point.z < params.camera.m_sensorDepthWorldMax && length(point - camPos) <= 0.04f) {
atomicMin(&minimum[warp], point.z*1000.0f);
}
}
@@ -457,7 +471,7 @@ __device__ inline float warpMin(float e) {
const float3 point = params.camera.kinectDepthToSkeleton(x+u, y+v, float(point_in.tex2D(x+u, y+v)) / 1000.0f);
// If it is close enough...
- if (point.z > params.camera.m_sensorDepthWorldMin && point.z < params.camera.m_sensorDepthWorldMax && length(point - minPos) <= 0.02f) {
+ if (point.z > params.camera.m_sensorDepthWorldMin && point.z < params.camera.m_sensorDepthWorldMax && length(point - minPos) <= 0.04f) {
// Append to neighbour list
//unsigned int idx = atomicInc(&nidx[warp], MAX_NEIGHBORS_2-1);
unsigned int idx = atomicAdd(&nidx[warp], 1);
@@ -469,9 +483,6 @@ __device__ inline float warpMin(float e) {
__syncwarp();
- // FIXME: What if minDepth fails energy test, an alternate min is needed.
- // Perhaps a second pass can be used?
-
const float maxDepth = float(maximum[warp])/1000.0f;
const float interval = (maxDepth - minDepth) / float(MAX_ITERATIONS);
@@ -506,8 +517,9 @@ __device__ inline float warpMin(float e) {
maxenergy = newenergy;
}
- // Search for first energy maximum above a threshold
- if (bestdepth > 0.0f && maxenergy >= 0.1f) {
+ // If enough energy was found and this thread was the one that found the best
+ // then output the depth that this energy occured at.
+ if (bestdepth > 0.0f && maxenergy >= ENERGY_THRESHOLD) {
//printf("E D %f %f\n", maxenergy, bestdepth);
const unsigned int cx = x;
const unsigned int cy = y;
@@ -517,6 +529,28 @@ __device__ inline float warpMin(float e) {
//depth(cx,cy) = bestdepth * 1000.0f;
}
}
+
+ // TODO: Could the threshold depend upon the number of points? Fewer points
+ // due to distance is incorrect since really there may not be fewer points
+ // Perhaps the best option is to make it depend on depth ... really close
+ // and really far both has lower thresholds due to point densities. Other
+ // option is smoothing factor and surface distances alter with distance to
+ // vary the number of points used ... smoothing factor could be a multiple
+ // of pixel size at given distance. Density from original source is also
+ // an influencer of smoothing factor and thresholds. Colour contrast also
+ // has a weighting influence, high contrast is high certainty in the
+ // disparity so such points should have a high influence over choice of
+ // surface location.
+ //
+ // Magnitude vs dispersion factor in the energy function ...
+ // * Mag is certainty of surface location
+ // * Dispersion is how far to propagate that certainty,
+ if (maxenergy >= ENERGY_THRESHOLD) return;
+
+ // Move to next possible surface...
+ clusterBase = minDepth + 0.04f;
+
+ };
}
// ===== Pass 2 and 3 : Attribute contributions ================================
--
GitLab