Skip to content

Releases: lamadipen/OpenFloodAI

v5.0.0

Choose a tag to compare

@lamadipen lamadipen released this 22 Sep 13:53
5c6e647

A new, redesigned console for setting up sites, downloading evidence, and reviewing results — built to replace scattered legacy pages with one clearer flow. Existing pages are untouched; this is an additional, separate UI.

Setting up a site

  • Create a site from a video — pick a video file when creating a site and its name, folder, and camera ID are filled in for you automatically (still fully editable).
  • See videos already on a site — the Setup tab now lists every video already added, with its purpose, dataset split, and whether it has a human label — no more guessing what's there.
  • More purpose options when adding a video — the full set of options (practice, possible rising/falling water, hard case, camera problem, etc.) is available, not just two.
  • Video ID fills in automatically — pick a video file on the Add Media form and its ID is derived from the filename, still editable.

Drawing the watched area and riverbank

  • Redraw the watched area from scratch — click and drag anywhere on the image to draw a brand-new watched area, or use "Clear and redraw" to start over, instead of only being able to nudge the existing box.
  • Works on video-only sites too — the watched-area and riverbank editors used to require a downloaded image sequence; they now also work directly on a site's video, letting you scrub to a frame and draw on it.
  • Multiple riverbank lines in one visit — trace as many banks as you need (e.g. left bank and right bank) in a single session; every line stays visible while you draw the next one, and each can be labeled, confirmed, or deleted independently.
  • See the watched area while tracing the riverbank — the watched area now shows as an outline while you draw riverbank lines, and new points are kept inside it, so you can't accidentally draw outside the area that's actually being analyzed.

Reviewing results

  • A new page for reviewing video results — browse every comparison window from a video validation run, see the machine's result and score on a clear chart, watch the exact video clip, and record what you see — all without leaving the page.
  • The watched area and riverbank show up on the video — a toggle overlays the saved watched area and riverbank lines directly on the video preview, so you can check they still line up with what the camera is showing.
  • Clearer, easier-to-read charts — the change-score chart now has gridlines, a trend line, and readable axis labels, and results use a consistent color scheme (green = no change, blue = cannot judge, red = possible change) throughout.
  • Always know what's set up and what isn't — the site overview page now clearly shows whether image-sequence review, video review, or both are ready to use, with a one-click way to run whichever one is missing.

Dataset tags

  • Tag individual videos, not just image sequences — assign a video to practice or locked-validation independently of any image sequence on the same site.

Gage (water-level) data

  • Gage data downloads automatically — when you download a timestamped image sequence for a recognized river camera, USGS gage readings for the same dates now come down with it, matched to each image's exact timestamp — no separate step required.
  • Optional — an "Also fetch USGS gage data" checkbox lets you skip this if you only want the images.

Getting around

  • Clearer navigation from the home page — the links to the console, review workspace, image downloader, and river camera tracker are now clear buttons with icons, not a cramped line of text.
  • A way back from the new console — the console now has a link back to the classic home page, so you're never stuck without a way back.

Behind the scenes

  • Several real bugs were found and fixed along the way: watched-area/riverbank points weren't being saved in the right coordinate system, a riverbank save was missing a required field, and video-only sites weren't correctly recognized by several pages.

v4.0.0

Choose a tag to compare

@lamadipen lamadipen released this 17 Sep 16:26
c1bab27

Highlights

  • Added a complete USGS timestamped image-sequence workflow with date selection, preview, sampling, download, and local review.
  • Added image-sequence validation with selectable baselines, water-level change results, review images, reports, and reproducible input snapshots.
  • Added a richer site details page for reviewing videos, configuration, validation history, evidence, and run comparisons.
  • Made human-label coverage clearer for each video and clarified the water-level label names.
  • Protected the shared watched area and riverbank guide when adding more videos, with an overlay preview to reveal camera-view mismatches.
  • Added image-sequence actions to both the guided workflow and classic site view.

All checks passed on the released main branch.

Full changes: v3.0.0...v4.0.0

v3.0.0

Choose a tag to compare

@github-actions github-actions released this 16 Sep 08:38
ac63fdc

What's New in v3.0.0

Building on v2.0.0's confirmed-riverbank-reference workflow, this release replaces that rectangle-based approach with a human-traced polyline guide that follows the real curve of a riverbank, and adds a new capability: pulling review video/image material directly from public archives and live camera streams into the Home UI.

1. Normal Waterline Guide — replaces the confirmed-reference rectangle (OF-087)

  • Trace one or more curved polylines marking the normal water edge, instead of a single straight-edged rectangle
  • Every line is its own row with an independent label, status (draft/confirmed/invalid), and notes; adding, deleting, or switching lines never discards in-progress edits
  • Bulk save (all lines in one request) and hard delete (no invalidation trail), alongside the existing single-guide save/invalidate
  • No more machine-suggested starting shape — the machine never proposes the trusted baseline; only a human draws it
  • The Watch viewer now shows every non-invalidated guide as soon as it's drawn (confirmed solid blue, unconfirmed dashed amber), not only once fully trusted
  • The Create Site setup form no longer offers guide drawing — guides are added afterward from their own panel
  • Fixed a layout bug where the Watch/Details View buttons could get clipped on narrower site cards

2. Download River Images and Video

  • New Home UI page for pulling review material straight from the public USGS HIVIS camera archive — no manual file transfer needed
  • Choose a camera, date, hour, and timezone to download up to four images across a chosen hour, with actual vs. requested capture times shown side by side
  • Download a camera's latest time-lapse video directly, independent of the image date/hour
  • Turn a batch of downloaded images into a local test video (five seconds per image) to exercise your site's intake, sampling, and labeling workflow end-to-end
  • All downloads are saved locally, kept separate from validation sites, and never auto-approved as training data

3. Live Camera Stream Clips — on demand and scheduled

  • The same downloader page can now save a short clip from any HTTPS live camera stream (an .m3u8 link), not just the USGS archive
  • Download a clip on demand with one click, or turn on a background schedule so the Home UI server keeps trying on an interval by itself — even with the browser closed
  • Built for a feed that isn't always reachable: a failed attempt cleans up completely and reports why, never saving a partial or corrupted file
  • Stream URLs are checked to reject anything pointing at a local or private network address

4. Real-World Grounding: Nepal Bhote Koshi 2026 Case Study

  • Added a summary of the Government of Nepal's assessment of the 26 August 2026 Rasuwa-Bhote Koshi outburst flood to the product/ML-readiness docs
  • Documents a real case where gauges failed and warning time was short, motivating why OpenFloodAI's validation work should preserve visual evidence of sudden events
  • Includes candidate future label/metadata ideas (event type, visual evidence, supporting evidence, system state) drawn from the report

All local-first, camera-based, and privacy-preserving as before: no cloud uploads, no alerts, no flood-risk claims. Archive and live-stream downloads only happen when you explicitly ask for them (or turn on a schedule yourself), and everything stays on your own computer.

Full changelog: v2.0.0...v3.0.0

v2.0.0

Choose a tag to compare

@github-actions github-actions released this 15 Sep 12:29
261adab

What's New in v2.0.0

Building on v1.0.0's core validation workflow, this release adds a full confirmed-riverbank-reference workflow: draw or suggest a trusted reference inside your watched area, see its quality checked automatically, and view it as a visual overlay on reports and video.

1. Confirmed Riverbank Reference (OF-082)

  • Draw a smaller, trusted "confirmed reference" rectangle inside the existing watched area
  • Draft / confirmed / invalid status, with an explicit invalidation reason (camera moved, view changed, bank changed, visibility unreliable)
  • Named stable markers (bridge pillars, rocks, wall edges) as extra reference points
  • Available both in the guided workflow and folded into the Create Site form

2. Riverbank/Reference Quality Checks (OF-083)

  • New per-sample quality fields on human labels: riverbank visible, stable marker visible, water boundary visible, camera stable, visibility condition
  • Automatic plain-language explanation of why a sample can't be trusted (e.g. "The camera does not appear stable in this sample.")
  • Validation reports now separate baseline-ready samples from practice-only ones

3. Riverbank Evidence in Validation Reports (OF-084)

  • Reports show the site's confirmed reference (video/time it was confirmed from) once per report
  • Per comparison window: whether that window's reference evidence was actually usable, and an honest note when the system can't tell water-coverage direction
  • Missing human label and reference-quality problems are shown independently — neither hides the other

4. Visual Overlays (OF-085)

  • The confirmed reference and its markers are now burned directly onto saved review images
  • New read-only "Watch" viewer: play any site video with the watched area, confirmed reference, and markers drawn live on top — never modifies the video file

5. Machine-Suggested Riverbank Guide (OF-086)

  • New "Suggest Reference" button proposes a draft riverbank rectangle using a simple local brightness-edge heuristic — no ML model, no segmentation
  • Clearly marked "Suggested — unconfirmed" until a human accepts it, adjusts it, or draws their own
  • Every confirmed reference now records whether it started as a suggestion or manual drawing

All local-first, camera-based, and privacy-preserving as before: no cloud uploads, no alerts, no flood-risk claims.

Full changelog: v1.0.0...v2.0.0

v1.0.0

Choose a tag to compare

@github-actions github-actions released this 10 Sep 22:37

Features Included In v1.0.0

  1. Standalone desktop app
    • Windows .exe
    • macOS .app zip
    • Runs without manually installing Python
    • Opens the local Home UI in browser
  2. Local Home UI dashboard
    • Shows all local sites
    • Shows site readiness
    • Shows missing setup items
    • Shows machine-ready vs human-comparison-ready status
  3. Guided workflow
    • Site setup
    • Video intake
    • Watched area
    • Human labels
    • Manifest
    • Run validation
    • Review results
  4. Create site from UI
    • Add site details
    • Add first video
    • Select watched area
    • Save config and manifest metadata
  5. Add video to existing site
    • Copy local video into site folder
    • Fill video metadata
    • Update manifest.jsonl
  6. Set watched area
    • Pick an existing site video
    • Move to a clear frame
    • Drag box around the area to monitor
    • Save reference_region
  7. Add human labels
    • Pick video
    • Set time window
    • Choose human label
    • Save label for comparison
  8. Run validation
    • Runs local video review
    • Creates machine records
    • Compares with human labels when available
    • Shows agree, disagree, or cannot_compare
  9. Run history
    • Each validation run gets its own folder
    • Old runs are preserved
    • Inputs used by the run are saved
  10. Review results in UI
    • View latest report
    • View report history
    • View review images
    • See plain-language result explanations
  11. Export results
    • Export one run as zip
    • Export all site reports
    • Raw video excluded by default
    • Uses saved run snapshots, not changed live files
  12. Delete local sites
    • Delete one selected site
    • Delete all local sites
    • Confirmation required
    • Deletes only local site folders