Project website: https://rayford295.github.io/drone-compression-hpc/
This repository studies high-resolution drone imagery compression as an HPC
workflow. It evaluates conditional diffusion compression (CDC), tiled diffusion
reconstruction, and downstream inspection on GPU supercomputers, with the final
SC26 submission files archived under sc26/.
The project does not propose a new compression algorithm. Instead, it treats CDC as a data-centric HPC workload: raw drone imagery creates storage and transfer pressure, while diffusion reconstruction creates GPU-memory and throughput pressure. The main question is which operating point makes full-resolution drone imagery practical for repeated analysis on NCSA DeltaAI GH200.
For the current Galveston drone-image workload, the best operating point is:
balanced checkpoint b00064 + 256 x 256 tiled reconstruction
This setting keeps full-resolution imagery usable while reducing the cost of reconstruction:
| Setting | Sec/img | Img/hr | Peak GPU | Compression | F1 | Roof IoU |
|---|---|---|---|---|---|---|
| High quality 512 | 88.51 | 40.7 | 2.96 GB | 29.73x | 0.8805 | 0.9001 |
| Balanced 256 | 79.34 | 45.4 | 1.57 GB | 79.98x | 0.8705 | 0.8989 |
| Max compression 256 | 78.91 | 45.6 | 1.57 GB | 139.89x | 0.8679 | 0.8956 |
Compared with full-image balanced reconstruction, balanced 256 tiling cuts peak GPU memory from about 52 GB to about 1.6 GB and reduces runtime from about 143 seconds to about 79 seconds per image.
The repository contains code, runbooks, and GitHub-safe result summaries for a measured systems study of CDC reconstruction on HPC hardware.
| Path | Purpose |
|---|---|
experiments/compression/ |
Compression, reconstruction, detection, SAM, and summarization workflows. |
results/ |
CSV tables, summaries, and small visual examples from completed runs. |
imgs/ |
Public-facing figure panels for posters and summaries. |
sc26/ |
Final PDFs uploaded for the SC26 submission. |
xparam/, epsilonparam/ |
CDC-derived model code paths used by the experiments. |
Selected figure panels from the SC26 submission are available in imgs/.
GH200 vs H200 hardware comparison — reconstruction throughput and memory behavior across the two GPU platforms:
Bottleneck proxy heatmap — where time is spent across compression settings and tile sizes:
Compute platform evidence scope — which claims are backed by measurements on which platform:
These figures are generated from committed result tables using:
python3 scripts/generate_sc26_figures.py
python3 scripts/generate_qualitative_visual_panel.pyMost readers should start with these result folders:
| Folder | Contents |
|---|---|
results/2026-06-05-tradeoff-n50-lpips/ |
Main N50 compression-setting x tile-size matrix with LPIPS. |
results/2026-06-12-yolo-vehicle-roof-human-n50/ |
Human-label vehicle and roof detection evaluation. |
results/2026-06-12-sam-vehicle-roof-human-n50/ |
SAM vehicle and roof mask-stability evaluation. |
results/2026-04-28-h200-reconstruction/ |
Delta H200 quick reconstruction comparison. |
results/2026-04-26-reconstruction/ |
Initial DeltaAI GH200 reconstruction profiling. |
See results/README.md for the full archive index.
This repository tracks source code, scripts, summary tables, and small GitHub-safe visual examples. It does not track raw drone-image collections, full-resolution reconstructed outputs, large model checkpoints, detector weights, local caches, or full HPC output directories.
The result tables are intended to make the systems story auditable without publishing private or oversized data artifacts.
If you use this repository, please cite the corresponding SC26 poster or project paper once available.