Skip to content

Latest commit

 

History

History
53 lines (53 loc) · 3.12 KB

TODO.md

File metadata and controls

53 lines (53 loc) · 3.12 KB

TODO list:

  • Add a clang-format style to the project
  • Add a clonable StreamPU task to compute the task gray2rgb in parallel
  • Add the --gnd-detect option to all of the fmdt-detect* exe binaries
  • Add a new criterion in k-NN: the distance between the bounding boxes (in addition to the distance between the centers of inertia)
  • Split tracking in two separated process: one for the tracking, the other for the classification
  • Use the --trk-ddev criterion in the motion registration to determine moving & non-moving objects
  • Write an interactive Python script to ease the human detection task (= ground truth)
  • In fmdt-log-parser, find the RoIs without the "tracks to RoIs" file.
  • Implement socket forward in AFF3CT-core will significantly speedup the CCL. More precisely, socket forward is required to wrap the features_labels_zero_init function in the runtime)
  • Implement --trk-ell-min in fmdt-detect-rt* (the classification with ellipse ratio)
  • Add CI tests with fmdt-check
  • Rewrite fmdt-check in Python
  • Regroup CCs together if they are close and if their velocity vectors are close too
  • Limit the detection to only a particular angular range (say +/- 30 deg from straight down)
  • Add zones to ignore for CCL (for instance, this can allow to do not take care of saturated zones)
  • For each detection: extract meteor video sequence
  • Add few frames before and after a meteor extraction : this way if we need to re-process it we don't have to run it through the whole video again
  • Compute extrapolated bounding boxes of tracks in fmdt-log-parser when generating the bounding boxes (with the --trk-bb-path option)
  • Put a saturation flag for each detected object: meteor and star -> this will help for photometry => in the connected-components analysis, compute the number of saturated pixels per RoI (if I_p = 255)
  • Support more image input formats
  • Add --video-loop and --video-buff support to the video module (based on ffmpeg-io)
  • Extrapolate more than one frame in tracking
  • Add memory check tests in the CI (valgrind --leak-check=full --show-leak-kinds=all)
  • Compute velocity of moving RoIs, add this to the statistics (no need to do this, this is the same thing as RoI error when is_moving = 1)
  • Tests column by column in the python script (new columns should not result in a failure test)
  • Use dynamic vector to store the tracks
  • Enable read from images in fmdt-visu
  • Document --out-mag output text file (objects magnitudes)
  • Add a parameter to select the number of threads to use in ffmpeg-io
  • Improve magnitudes management in the tracking (keep array sizes per objects + memory reallocations => like in C++ std::vector. What about creating a vector module in C for this? It will also be useful for bounding boxes)
  • Add classification reason (from meteor to noise) in the tracks output
  • Implement features_compute_magnitude with FMDT multi-threaded runtime