Change Detection leads with the change itself — even objects too small to name.

• AddedChange hotspots · scene comparison. Change Detection now surfaces the changed regions directly as a clean before/after view, independent of whether any model can classify the object. This catches things that are genuinely changed but below the recognition limit at drone altitude — e.g. a chair placed on the apron shows up as a high-confidence "added object" with a side-by-side before/after crop, so the analyst sees it even though no detector can label a top-down chair.
• ImprovedSmall-object sensitivity. The hotspot pass ranks changes by strength (not just size) and keeps compact objects that the standard size filter would discard, while separating real placed objects from hangar-roof parallax.
• ImprovedThe scene comparison is anchored on a detected aircraft for a reliable alignment between the two flights, and shown at the top of the change result.

Change Detection now catches arriving/departing aircraft & clearer before/after evidence.

• FixedA newly-arrived aircraft is now reported. A small aircraft that is on the ground for only a few seconds of a flight is a fast, transient target the pixel-level change step cannot track into a stable region — so it was being missed. Change Detection now also runs a full object-inventory comparison between the two flights: anything the detector finds in one flight but not the other (e.g. aircraft 0 → 1) is surfaced as added or removed, with the object's image — matching the Object-Detection page.
• AddedFlight A → B object inventory. A per-category tally (personnel · vehicle · vessel · aircraft) for each flight with the delta, so an analyst sees at a glance what the AI counted in each pass and what moved between them.
• ImprovedClean before/after evidence. The "what changed" table now shows un-annotated side-by-side before (A) and after (B) crops of each object — no boxes drawn over it — so the difference reads clearly.

Console refinements — live processing speed & a tidier workspace.

• AddedReal-time processing speed. Every analysis now shows how long it took on the RTX 4090 and its throughput (frames per second), with a live elapsed timer while a job runs — so the hardware's speed is visible at a glance.
• ImprovedCollapsible activity panel. The live job panel (bottom-left) can be minimized to a compact badge and re-opened with one click; it shows a running-job count and remembers your preference. Keeps the screen clean during a demo or presentation.
• ImprovedGeneral interface polish across the results header and job list for clearer, faster reading.

Infrastructure upgrade (2× RTX 4090) & object-detection recalibration.

• ImprovedMigrated to a dual RTX 4090 server. The platform now runs on a far more powerful GPU host (24 GB RTX 4090), pinned to a dedicated GPU, so object-detection and change-detection jobs complete materially faster. Same URL, same login.
• FixedObject detection was over-sensitive. It mislabelled runway holes and road-line markings as "pedestrian" and inflated counts. Recalibrated with per-category confidence floors (high bar for personnel), so those false positives are eliminated — an empty field went from ~125 phantom detections to near zero.
• FixedObject counts no longer inflated. A moving drone made the simple tracker split one object into many; added camera-motion-compensated tracking plus track-merging, cutting a sample clip from 67 "objects" to ~9 that match what's actually on the ground (≈5 cars + 1 aircraft).
• AddedAircraft detection. A dedicated aerial (DOTA) detector now runs alongside the VisDrone model and is fused in, so a small fixed-wing plane is correctly labelled aircraft instead of "bus".
• FixedSingle aircraft no longer double-counted. A painted closed-runway "X" marking was being detected as a second plane; a colour check (real aircraft are coloured; paint is achromatic white-on-grey) now rejects it, so one plane reads as one aircraft.
• AddedObject thumbnails in the tables. Every row in the object-detection tracked objects table and the change-detection what changed table now shows a cropped image of the actual object, so an analyst can verify each detection at a glance.

Aerial object detection overhaul & object-level change attribution.

• FixedObject detection now finds top-down vehicles. The generic (COCO) detector could not recognise pure overhead views — a parked car/van read as "couch" or nothing. Replaced the default with a VisDrone-trained aerial model built for low-altitude drone imagery (car · van · truck · bus · pedestrian). On a real patrol clip this lifted the count from 2 false detections to 16 real objects (14 cars at up to 0.92 confidence).
• ImprovedTiled (sliced) inference. Each frame is split into overlapping tiles analysed at near-native resolution, so a tiny ~80 px car becomes large enough to detect reliably. Memory-safe on the on-prem 6 GB GPU.
• ImprovedPer-category confidence floors & tracking. Vehicles/vessels/aircraft stay sensitive while the noisiest class (aerial "pedestrian" over textured ground) must clear a high bar — an empty field dropped from ~125 false hits to ~5. Each object is tracked across frames and counted once.
• AddedChange Detection now reports what changed. Every change region is cross-examined with the aerial detector on the before & after frame and classified Added · Removed · Modified · Unknown, with the object identified (e.g. "added car"). The report leads with the tally and a "What changed" table; changes.json carries the change type and object class/category.
• Added"High-sensitivity tiled scan" toggle in the object-detection controls, and this Changelog page.
• ImprovedTechnical whitepaper updated to document the aerial model, tiled inference and object-level change attribution.

Real-flight alignment, footage inspection & faster loading.

• FixedRobust two-flight image alignment. Real back-to-back DJI passes were badly misaligned — an over-fitting homography warped one frame off-screen and flooded the heatmap. Rebuilt registration on SIFT features + a stable similarity/affine model with geometric, ≥70% overlap and photometric validation, plus no-data masking. Mis-paired frames are now rejected instead of producing false changes (real A→B dropped 215 → 123 changes, zero black warps).
• AddedInspect the source clips on the web. Every result shows the input footage next to the AI output — before/after side-by-side for change detection, raw vs annotated for object detection — plus a "Preview source clips" button. HTTP range streaming so large clips seek instantly.
• AddedExpanded evaluation set — a 19-scenario, 42-change synthetic benchmark spanning naval harbours, the border and riverine Bangkok, alongside three real DJI passes over one route.
• ImprovedFast page loads. Scenario thumbnails served as cached, resized JPEGs (31 MB → 91 KB) and whitepaper figures recompressed and lazy-loaded.

Object detection module & technical whitepaper.

• AddedMilitary object detection. Inventory every object of interest in a single clip — personnel · vehicle · vessel · aircraft — with per-object tracking and an annotated playback video, order-of-battle summary and machine-readable objects.json.
• AddedTechnical whitepaper covering both modules, accuracy and how it is evaluated, supported object types, security & deployment, and the API.

Initial deployment.

• NewChange Detection platform live at drone.iapp.co.th — compare two flights over one route and surface what changed, with a ranked report, gallery and a 4-panel overlay video.
• NewOn-prem deployment: GPU-accelerated pipeline behind a password-protected HTTPS console, designed to run entirely on the unit's own machines — no data leaves the building.