Marc Freir

SSL F3 Results

SSL Experiments on the F3 Seismic Dataset

This page collects all interactive visualizations and summary tables generated from the augmentation study on the F3 seismic dataset, comparing seven self-supervised learning (SSL) methods across four augmentation phases.

Methods: BYOL · DenseCL · FastSiam · KenshoDense · MoCo · SimCLR · TriBYOL

Metric: mean Intersection over Union (mIoU) · Phases: Phase 1 (parameter optimization) · Phase 2 (pairwise transforms) · Phase 3 (triplets) · Phase 4 (random apply)

Each scatter point in the pairwise plots represents one augmentation configuration (averaged over runs), colored by augmentation phase. Dotted reference lines indicate the paper-original baseline where available.

All Plots — Combined Page Start here

Single self-contained HTML with every figure — SPLOM, 21-pair grid, and all 21 individual pair plots — in one scrollable page with a sticky navigation sidebar.

f3_all_plots_combined.html

The full report in one page: SPLOM, 4×6 pair grid, and all 21 individual scatter plots with baseline reference lines and phase-colored points. Includes a dark sticky sidebar for quick navigation between figures.

23 figures · ~50 MB · opens in new tab

SPLOM & Overview Figures TriBYOL-centric

Multidimensional and grid views showing TriBYOL against all other methods simultaneously.

Scatter Plot Matrix (SPLOM) — all 7 methods in a single 7×7 interactive matrix. Each cell shows one pair; axes are avg mIoU per transformation. Only transforms common to all methods are included. Points colored by augmentation phase (1–4).

7×7 matrix · 21 pairs at once · hover for values

f3_pairwise_grid.html

4×6 subplot grid — all 21 individual pair scatter plots arranged in a single panel. Each subplot has the diagonal y = x reference line and dotted baseline lines. Useful for a quick side-by-side overview of all comparisons.

21 subplots · 4 rows × 6 cols · ~wide screen recommended

All 21 Pairwise Comparisons Individual

Each plot shows avg mIoU per transformation for one pair of methods. The X-axis method is listed first. Dotted reference lines mark paper-original baselines where available.

TriBYOL vs …

TriBYOL vs BYOL

Compares TriBYOL (X) against its predecessor BYOL (Y). Points above the diagonal indicate configurations where BYOL outperforms TriBYOL, and vice versa. Both paper baselines are shown as dotted lines.

Baselines available for both methods

TriBYOL vs KenshoDense

TriBYOL (X) against KenshoDense (Y), a dense SSL method using multiple views. Reveals which augmentation strategies benefit dense feature learning.

Phase-colored · diagonal reference

TriBYOL vs FastSiam

TriBYOL (X) vs FastSiam (Y). FastSiam uses a fast Siamese architecture; this plot highlights augmentation regimes where each approach has an edge.

Phase-colored · diagonal reference

TriBYOL vs DenseCL

TriBYOL (X) vs DenseCL (Y), a contrastive dense SSL method. Useful for understanding how bootstrapping (TriBYOL) compares to contrastive objectives on seismic data.

Phase-colored · diagonal reference

TriBYOL vs MoCo

TriBYOL (X) vs MoCo (Y), a momentum-based contrastive method. Shows how momentum contrast differs from bootstrapping across different augmentation phases.

Phase-colored · diagonal reference

TriBYOL vs SimCLR

TriBYOL (X) vs SimCLR (Y). SimCLR baseline is available for reference. A key comparison since SimCLR uses natural-image augmentations that may not transfer well to seismic data.

Baselines available for both methods

BYOL vs …

BYOL vs DenseCL

Bootstrap (BYOL) vs contrastive dense (DenseCL). Both benefit from seismic-specific transforms.

BYOL baseline available

BYOL vs FastSiam

Two bootstrapping-family methods compared across all augmentation phases.

BYOL baseline available

BYOL vs KenshoDense

Standard BYOL vs the dense multi-view KenshoDense variant on seismic features.

BYOL baseline available

Bootstrapping vs momentum contrast — highlights how augmentation phase interacts with the SSL objective.

BYOL baseline available

Two landmark SSL methods, both with paper baselines available as dotted reference lines.

Baselines available for both methods

DenseCL vs …

DenseCL vs FastSiam

Dense contrastive learning vs fast Siamese networks on seismic augmentations.

Phase-colored · diagonal reference

DenseCL vs KenshoDense

Two dense SSL methods — contrastive (DenseCL) vs multi-view bootstrapping (KenshoDense).

Phase-colored · diagonal reference

DenseCL vs MoCo

Dense vs standard momentum contrast, revealing whether dense features help on F3.

Phase-colored · diagonal reference

DenseCL vs SimCLR

Dense contrastive vs standard SimCLR. SimCLR baseline available as reference.

SimCLR baseline available

FastSiam vs …

FastSiam vs KenshoDense

Two multi-view bootstrapping methods with different architectural choices.

Phase-colored · diagonal reference

FastSiam vs MoCo

Siamese vs momentum-contrast; shows how negative-free methods fare against contrastive ones on seismic data.

Phase-colored · diagonal reference

FastSiam vs SimCLR

Negative-free Siamese vs contrastive SimCLR. SimCLR baseline available.

SimCLR baseline available

KenshoDense vs …

KenshoDense vs MoCo

Multi-view dense bootstrapping vs momentum contrast — highlights the role of view count and dense features.

Phase-colored · diagonal reference

KenshoDense vs SimCLR

KenshoDense's dense multi-view approach vs SimCLR's standard contrastive objective. SimCLR baseline available.

SimCLR baseline available

MoCo vs …

Two classical contrastive SSL methods. Both use a memory queue / negative pairs; SimCLR baseline is available as reference.

SimCLR baseline available

Individual Pair Files Directory

All 21 individual HTML files also live in the f3_pairwise_individual/ directory. Each file is a fully standalone Plotly figure (no internet needed after loading) and can be shared independently or embedded in slides. The naming convention is:

f3_pairwise_individual/
├── f3_pair_TriBYOL_vs_BYOL.html
├── f3_pair_TriBYOL_vs_KenshoDense.html
├── f3_pair_TriBYOL_vs_FastSiam.html
├── f3_pair_TriBYOL_vs_DenseCL.html
├── f3_pair_TriBYOL_vs_MoCo.html
├── f3_pair_TriBYOL_vs_SimCLR.html
├── f3_pair_BYOL_vs_DenseCL.html
├── f3_pair_BYOL_vs_FastSiam.html
├── f3_pair_BYOL_vs_KenshoDense.html
├── f3_pair_BYOL_vs_MoCo.html
├── f3_pair_BYOL_vs_SimCLR.html
├── f3_pair_DenseCL_vs_FastSiam.html
├── f3_pair_DenseCL_vs_KenshoDense.html
├── f3_pair_DenseCL_vs_MoCo.html
├── f3_pair_DenseCL_vs_SimCLR.html
├── f3_pair_FastSiam_vs_KenshoDense.html
├── f3_pair_FastSiam_vs_MoCo.html
├── f3_pair_FastSiam_vs_SimCLR.html
├── f3_pair_KenshoDense_vs_MoCo.html
├── f3_pair_KenshoDense_vs_SimCLR.html
└── f3_pair_MoCo_vs_SimCLR.html

Summary Charts Overview

High-level bar charts for a quick comparison of all methods.

f3_overall_summary_bar.html

Overall summary bar chart — average mIoU ± std across all runs for each method, with max-value diamond markers and horizontal dotted lines for paper-original baselines (BYOL, SimCLR, TriBYOL).

7 methods · mean ± std · baselines

f3_phase_avg_bar.html

Per-phase grouped bar chart — average mIoU per augmentation phase (1–4) per method, with error bars showing std. Useful for identifying which phase benefited each method most.

4 phases × 7 methods · grouped bars

f3_heatmap_methods_x_transforms.html

Methods × Transformations heatmap — RdYlGn color scale showing mIoU for each (method, augmentation config) pair. Top 60 transformations by average mIoU across all methods. Good for spotting which transforms generalize across methods.

7 methods × top-60 transforms · hover for values

Result Tables (CSV) Data

Machine-readable CSV files with all aggregated results. Open in Excel, LibreOffice, or load with pandas.read_csv().

f3_summary_table.csv

Overall statistics per method: number of runs, number of transforms, mean / std / max / min mIoU across all runs, transform-level averages, and paper-original baseline values where available.

7 rows · one row per method

f3_per_phase_table.csv

Per-method per-phase breakdown: mean, std, and max mIoU for each of the four augmentation phases (Phase 1–4) for all 7 methods.

28 rows · 7 methods × 4 phases

f3_common_transform_table.csv

Per-transformation breakdown for configurations present in all 7 methods. Columns: transform name, phase, mIoU per method, mean and std across methods. Sorted by average mIoU descending — top rows are the augmentations that worked best across all methods.

One row per common transform · sorted by mean mIoU

All projects