wenmin-wu/cv-gaussian-sphere-target-generation

Gaussian Sphere Target Generation

Overview

When ground truth is point annotations (x, y, z) rather than voxel masks, generate training targets by placing 3D Gaussian blobs at each annotated location. Each particle type gets its own channel with a configurable radius. The model learns to predict these soft targets, and centroids are recovered at inference via peak detection or connected components.

Quick Start

import numpy as np

def generate_gaussian_volume(shape, coords, sigma=3.0):
    volume = np.zeros(shape, dtype=np.float32)
    for z, y, x in coords:
        zz, yy, xx = np.ogrid[
            max(0,int(z)-3*int(sigma)):min(shape[0],int(z)+3*int(sigma)+1),
            max(0,int(y)-3*int(sigma)):min(shape[1],int(y)+3*int(sigma)+1),
            max(0,int(x)-3*int(sigma)):min(shape[2],int(x)+3*int(sigma)+1),
        ]
        d2 = (zz-z)**2 + (yy-y)**2 + (xx-x)**2
        volume[zz, yy, xx] = np.maximum(volume[zz, yy, xx], np.exp(-d2/(2*sigma**2)))
    return volume

Workflow

1. Parse point annotations (z, y, x) per particle class
2. For each class, create a zero volume matching the input shape
3. Place Gaussian blob at each coordinate (clip to volume bounds)
4. Use np.maximum to handle overlapping particles (keep brightest value)
5. Stack per-class volumes into multi-channel target tensor

Key Decisions

• Sigma: Match to expected particle radius; too large → merged blobs, too small → hard to learn
• Per-class channels: Separate channel per particle type enables multi-class detection
• Max vs sum: max prevents double-counting overlapping particles; sum better for density estimation
• Truncation: 3-sigma cutoff balances accuracy vs speed

References

• DeepFindET_Train