Background subtraction is a common and widely used technique for generating a foreground mask by using static cameras. In controlled environments, motion detection can be separated from static background using statistical modeling. Instead of running a full object detector, we model the background over time and isolate dynamic foreground regions.
TECHNIQUES
Frame differencing
Running average background
Gaussian Mixture Models (MOG2)
Lightweight
Works without deep learning
The Goal of interactive segmentation is to cut out the foreground object in photo, based on some input from the user. More precisely, we aim to associate a binary level to each of the pixels in the image, indicating whether this pixel belongs to the foreground or background, based on the observer RGB data at each pixel. Automatic object detection from Meta's SAM is revolutionized segmentation because this model was trained on billions of masks. Segmentation can be used with a point click, multiple clicks, drawing a bounding box or a text prompt.
AREAS OF INTEREST
Product isolation
Medical imaging
Industrial inspection
Dataset preparation
Manual bounding boxes
Iterative refinement
Stereo vision systems provide depth perception and spatial information enabling to compute the distance between objects based on their disparity. Stereo vision estimates depth (3D structure).
AREAS OF INTEREST
Robotic grasping
Robotic grasping
Bin picking
Autonomous navigation
Pallet dimension measurement
volume estimation
3D estimation
Depending on the application, this may involve: changing image dimensions, relocating objects, removing objects, filling missing regions. Unlike simple pixel manipulation, Random Field approaches model spatial dependencies between neighboring pixels. This ensures that textures, edges, and structures remain coherent after transformation.
AREAS OF INTEREST
Image stitching
Industrial part alignment
Object replacement
Augmented reality overlays
Document correction
Template matching
Texture synthesis focuses on learning a generative model from a small sample of texture so that newly generated samples appear as natural extensions of the original pattern. By preserving these local relationships, the generated texture maintains visual consistency.
AREAS OF INTEREST
Surface and material simulation
Synthetic dataset generation
Texture-based defect modeling
Pattern expansion
The Laplacian of Gaussian (LoG) is a classical edge and blob detection technique that highlights regions where image intensity changes rapidly. It works by first smoothing the image with a Gaussian filter (to reduce noise), then applying the Laplacian operator (second derivative) to detect zero-crossings that indicate edges or blob-like structures.
AREAS OF INTEREST
Zero-crossing detection
Multi-scale space detection
Noise sensitivity
Smoothing necessity
Edge smoothness
Classical machine vision method and reliable feature extraction before deep learning. Canny Edge Detection identifies strong structural boundaries using gradient analysis, non-maximum suppression, and hysteresis thresholding — producing clean, connected edges. Harris Corner Detection identifies stable interest points where intensity changes in multiple directions — enabling tracking, motion estimation, and geometric reconstruction.
TECHNIQUES
Industrial inspection
Structure zoom-in
Zero-crossing detection
Gradient magnitude & direction
Structure tensor
Corner response function
