MEDiC: Evolved Part Masking Visualization

Interactive demo of Evolved Part Masking from MEDiC. Uses a frozen CLIP ViT-B/16 for attention extraction + EM/HC clustering for semantic part discovery. Alpha and K auto-adjust as you move the epoch slider, simulating how masking evolves during training.

Input Image

Masking

Mask Ratio

0.1 0.9

Clustering

Algorithm

EM (Expectation-Maximization) Hierarchical Clustering (HC)

K min

3 30

K max

10 50

EM Iterations

5 30

Spatial Noise

Type

Grid Random

Training Schedule

Gamma (0.5=sqrt, 1.0=linear, 2.0=quadratic)

0.1 3

Current Epoch

0 300

Total Epochs

100 500

Auto-computed (from epoch, gamma, K range)

Alpha

K range

Random Seed

0 100

Configuration

Masking Visualization

Individual Semantic Parts

Examples

How it works

Alpha controls the spatial-to-semantic transition: alpha = ((epoch+1) / total_epochs)^gamma

Early training (alpha near 0): grid-based spatial masking
Late training (alpha near 1): semantic part-based masking

K (cluster count) decreases during training from K_max to K_min, making parts coarser over time.

Gamma controls the evolution speed: 0.5 = fast start (sqrt), 1.0 = linear, 2.0 = slow start (quadratic)

Ref: "Evolved Part Masking for Self-Supervised Learning" (CVPR 2023) integrated into MEDiC.