GitLab

Improved confusion matrix generation, so it uses samples that are roughly uniformly distributed between the 2 classes, across the entire validation set. The matrix agrees with the reported accuracy, suggesting something is wrong with my implementation used with OpenCV in c++ (as the acc. is lower).

Fixed DNN detection implementation (now works fully), and refectored some paths to work in development. Reduced epochs to 14, as suggested by training data. Need to fix how app reads files that are in the process of being written.

Added detection logic for TensorFlow dnn model into cpp. Created more ground truth data. Modified dnn model for version 3, during-epoch data shows high accuracy, but have yet to finish training.

Added base training module for an Inception v3 architecture model; to transfer to IDUN with dataset and train.

Modified some relative paths to be compatible with final binaries/executable, that can now be created with pyinstaller. Also added binaries to gitignore.

All GUI elements finished, with proper initialization, handling, thread-execution and safety, termination, and serialization. Now works as a standalone software product. (Stopping while trimming a video means that thread waits to finish writing a detection before exiting)

Extended main.cpp for future selection of detection algorithm, now allowing adding different methods easily via an enum switch.

Parameter input (input folder, file prefix) in GUI operational, with serialization. Some inconsistency of moving data between GUI scenes. Not yet hooked up to core detection and trim functionality.

Finished final output reporting, according to spec. Currently uses Windows creation time to label outputs; might swap to file metadata. Working on gui next.

SVM with HoG features implemented, seems to be working well on single tests, but model needs to be serialized and connected to timestamp generator.

Added functionality to dataset generator; now works with multiple videos, and extracts a given percentage per video.

Restructured for python utilities package, and made standalone utility to generate labelled dataset that can easily be fed into keras later.

Fixed .csv reading, now correctly finds ground truth in input folder if it exists. Changed notation to 'groundTruth_[file]'

Fully functional Jaccard index/similary score complete. Will now look for ground truth file at end of computation, with name groundTruth[videoFile].csv, and calculate score. Note that the trimming functionality is commented out for faster testing of model.

Connected the Python parts that retrieve the generated .csv and trim the video based on the timestamps. Haven't been able to test, but it should work as a complete MVP. Have yet to add the optional success rate calculation functionality if there is a ground truth to compare to.

Merge branch 'addCuttingVideo' of https://git.gvk.idi.ntnu.no/LeoDLG/nina-thesis into addCuttingVideo