{"name": "Bambi - QGIS Integration", "package_name": "bambi_wildlife_detection", "version": "9.0.0", "experimental": true, "qgis_min": "3.22.0", "qgis_max": "3.99.0", "downloads": 26, "uploaded_by": "cpraschl", "upload_datetime": "2026-11-12T05:01:39.336287", "changelog": "Changed (breaking):\r\n    - One camera model for the whole plugin. Geo-referencing, FoV polygons, the inspectors and the labelling tool already used the corrected pose of core.camera_pose; ALFS mosaics, the per-frame GeoTIFF export (and the orthomosaic built from it) and the Correction Wizard's light-field preview now render with the same pose instead of their own. Rendered frames therefore land exactly where detections and tracks are geo-referenced.\r\n    - What changes for rendered outputs: (1) tilted (non-nadir) frames - the heading now turns about world up, as for geo-referencing; before, ALFS and GeoTIFFs were up to ~30 degrees off for oblique frames. (2) Translation corrections act along east / north / up, as for geo-referencing; before, rendering turned them with the camera heading. (3) The x rotation correction has the same sign as in geo-referencing. (4) The GeoTIFF export no longer doubles the rotation correction for the frame footprint. Nadir flights without corrections render as before.\r\n    - Migration: ALFS mosaics, GeoTIFFs and orthomosaics of flights with correction factors or tilted frames should be re-run. Correction values found with the wizard's two-click solver keep their meaning (it already used the geo-referencing camera); values tuned by eye in the wizard's ALFS preview of an earlier version may need re-checking - the 3D Flight Viewer's Corrections tab shows the effect live. External tools that consumed corrections in the old rendering convention (e.g. the TRex connector note about 2x rotations) are not affected by this plugin change but no longer match its renders.\r\n    - Every flight needs a name given by the user, also in a single-flight project. In the Flight box on the Input tab a new flight is added by typing its name and pressing \"+\" next to the Name field (names are unique; \"Thermal\", \"RGB\" and \"BAMBI Survey\" are reserved); \"+\" no longer opens a folder dialog - the flight's target folder is set in the Target folder field afterwards, which now makes the folder checks (a folder used by another flight is refused instead of switching to that flight; a folder processed before is adopted with its stored configuration after a confirmation). \"Browse and Add...\" adds an existing flight by its target folder. Setting a target folder before any flight exists uses the Name field or asks for the name, and no pipeline step runs for an unnamed flight. Projects saved by an earlier version keep their folder-derived name.\r\n    - Layer tree: inside a flight's group the layers are collected per camera in a \"Thermal\" and an \"RGB\" subgroup, and the plugin's layers keep one fixed order regardless of the order they were added in (top to bottom: survey area and exclusion zones, flight route, FoV, density heatmap, coverage map, transect areas, track inventory, matched pairs, tracks, detections, distance sampling, segmentation masks, reconstruction rasters, orthomosaic, ALFS, per-frame GeoTIFFs). Existing layers are never moved, so a manual arrangement and the user's own layers stay as they are. Scripts or styles that looked up a product group directly below the flight group find it one level deeper now.\r\n    - One camera selector per stage instead of one per step: \"Camera\" at the top of the Pre-Processing steps (extraction, flight route, FoV, ALFS, GeoTIFFs, orthomosaic), \"Camera\" at the top of Detection and Tracking (detection, geo-referencing, perpendicular distances, tracking) and \"Input\" (Thermal / RGB / Matched) at the top of Classification (embeddings and all classifiers). Steps of one stage can no longer be set to different cameras - in particular the classifiers no longer take different inputs; a project that stored different cameras per step loads with the stage's first step (species for classification) deciding. The export keeps its own camera selector.\r\n    - Duplicated (frozen) thermal frames are kept and flagged instead of dropped (Configuration > Extraction > \"Keep duplicated thermal frames\", on by default; needs bambi_detection 1.2.1 and alfs_py 3.1.0, which the Dependency Manager now installs). While a thermal camera's shutter is closed for calibration the video repeats the previous image; these frames now stay in poses_t.json marked \"duplicated\", so the frame index equals the video's frame number (one video, no skip or sampling) - which TRex tracklets rely on. The frame list is never shortened; image-based steps skip flagged frames (detection, frame and track matching, FoV, ALFS, GeoTIFFs, segmentation, embeddings, COCO / YOLO / Camtrap DP export, label boxes, Correction Wizard, and trail / live ALFS / auto-refine / thermal texture in the 3D Flight Viewer), pose-based ones keep them (flight route, transects, height profile) and the tools show a \"frozen frame\" note. Breaking: re-extracting a flight that was extracted before 9.0.0 changes its thermal frame numbers - the plugin asks first and marks everything computed from the old frames out of date; labels and frame-range corrections must be checked by hand. The TRex import warns when the frames were extracted with duplicates dropped.\r\n    Added:\r\n    - Custom implementations for detection, tracking and every classification stage: a class from your own .py file replaces the built-in method (Configuration > Detection / Tracking, and \"Custom implementation...\" as a model source in the Classification table and per species). setup() declares the class names, the supported cameras (thermal / RGB / matched) and - for classifiers - whether it reads the DINOv3 embeddings or the cut-out boxes (no embedding step needed) and whether it decides per frame (the plugin votes per animal) or per track. Crops follow the Classification tab's Projection (perspective frames or orthorectified GeoTIFFs). A tracker says whether interpolating the gaps of its tracks is applicable, with a tick box to switch it on or off. Every kind is given the camera extrinsics of its frames with the correction factors applied (position in mesh-local and project coordinates, rotation, field of view), and classifiers also each row's pixel and geo-referenced box - e.g. for geo-fencing. Trackers are called frame by frame (update(), with optional image loading and a final finish()) or once with the whole flight (run()). Classifiers get a flight in batches and embeddings are read on demand, so memory does not grow with the flight; the progress bar moves per frame, batch and animal. Free-form JSON options, a \"Check\" button that tests an implementation in the background step by step (declared packages, import, setup, test run) with a copyable log, validation of everything a class returns, and a confirmation before a file runs for the first time on a computer or after it changed. \"Download sample implementations...\" provides the built-in detector and classifier as custom classes, Microsoft's Overhead Wildlife Locator (OWL) and DroneMegaDetector as alternative detectors (not run against the real models by the authors), an IoU tracker, a crop-based and a track-level classifier. The dependencies of custom code are not managed by the plugin (install them in the OSGeo4W Shell / QGIS' Python). Note that installing or upgrading packages for custom code can change ones QGIS, the plugin or other plugins rely on and break steps that worked before. See docs/custom-implementations.md.\r\n    - Three kinds of imagery for detection and classification: Configuration > Detection > Projection chooses between the camera frames (as before), the per-frame GeoTIFFs and the ALFS light-field integrals of sampling mode, for the built-in and custom detectors alike. Boxes found on a GeoTIFF or an ALFS integral are placed in their camera frame through the raster's georeference, the DEM and the frame's corrected camera, so everything after detection works unchanged. The Classification Projection gains ALFS as well (crops of the embedding step and of custom classifiers; no default classifiers for ALFS imagery yet). This prepares ALFS detectors and classifiers.\r\n    - Survey Orthomosaic (Survey Analytics tab): one orthomosaic over every ticked flight - the flight's Orthomosaic step across flights, with the same merge of the exported frame GeoTIFFs and the merge mode of Configuration > Orthomosaic. Flights are merged in the order of the flight list; written to the survey's analytics folder and added to the BAMBI Survey layer group.\r\n    - 3D Flight Viewer (toolbar and plugin menu): a separate window with the active flight in 3D - the DEM mesh, the flight route and every camera pose with its correction factors applied, navigable by orbit, pan and zoom.\r\n    - Replay: play the flight in real time (from the frame timestamps) or frame by frame; each frame is draped onto the DEM by projective texture mapping, optionally keeping a trail of all frames shown as a continuous orthophoto. RGB or thermal camera, the current frustum and its ground footprint, and a follow-camera mode that locks the view to the drone camera - without it the view moves freely while the replay runs.\r\n    - Annotations in 3D: the frame's bounding boxes draped onto the DEM, or tracks with their current box and their history as a line on the ground.\r\n    - Live ALFS: the integral of the N frames around the current frame (N/2 before to N/2 after, with stride), rendered in the background and draped onto the DEM - fixed top-down or from the on-screen camera - with a focus offset above the DEM, a minimum number of frames per pixel, and export as a georeferenced PNG.\r\n    - Correction factors in the viewer, next to the Correction Wizard: edit the default and frame-range corrections and see route, frames and ALFS follow; compare two distant frames; an ALFS sharpness score with history; Auto-refine searches the factors that give the sharpest integrals. Saving writes correction.json (with a backup), updates the panel and can mark pose-based results as out of date.\r\n    - 3D reconstruction from the RGB frames with the DEM as prior: sparse points triangulated with the fixed poses, a dense surface from pairwise stereo (height of vegetation and objects above the bare-earth DEM), textures from the RGB or the time-linked thermal frames with occlusion handling. Written as PLY, COLMAP model, DSM / height-above-DEM / texture GeoTIFFs and a surface mesh that can serve as ALFS focal surface.\r\n    - 3D Flight Viewer annotations drawn in 3D on the DEM with adjustable line width: current boxes solid, boxes seen earlier dashed (overview of what has been seen); click a camera position to jump to its frame.\r\n    - 3D Flight Viewer options: hide the black undistortion border of the frames with the camera's mask, eroded at its edge like the GeoTIFF export (draping, live ALFS, auto-refine, reconstruction), line width and camera marker size.\r\n    - 3D Flight Viewer trail as a continuous orthophoto: every frame shown is painted once into one top-down texture that stays on the DEM, like a scanner sliding along the route.\r\n    - 3D Flight Viewer \"Show orthomosaic\": drapes the flight's orthomosaic onto the DEM as a static alternative to the trail.\r\n    - Reconstruction quality: the dense surface comes from pairwise stereo between neighbouring keyframes (each pair oriented from its own images and scaled on the DEM's ground, merged by a median vote per ground cell), a high-resolution orthophoto texture is draped on it, and an optional step refines the keyframe poses (and the field of view) from the images before reconstructing.\r\n    - AGL Profile: with a reconstruction from the 3D Flight Viewer, the reconstructed surface (DSM) is drawn above the DEM (DTM) with the camera's height above it (checkbox); the reconstruction also writes every camera position's nadir distance to surface and DEM as CSV.\r\n    - DEM detail in the 3D Flight Viewer (High / Medium / Low = all, 66 % or 33 % of the triangles) for faster draping and live ALFS on large DEMs, with the resulting height error shown; ray casting and all plugin products keep the full DEM.\r\n    - Dependency Manager: optional \"3D Flight Viewer\" group with separate install buttons for PyOpenGL 3.1.10 and pyqtgraph 0.13.7.\r\n    Fixed:\r\n    - Dependency Manager: a CUDA build of the newest tested PyTorch release (e.g. torch 2.11.0+cu126, torchvision 0.26.0+cu126) was shown as \"not tested\". The local build tag (+cu126, +cpu) is not part of the release number, but was compared as if it made the version newer than the tested maximum.", "external_deps": null, "download_url": "https://plugins.qgis.org/plugins/bambi_wildlife_detection/version/9.0.0/download/"}