Computer vision mini-projects using Airctic/ IceVision
Focus: Starting with IceVision, pets dataset, mmdet, retinanet/resnet, bboxes, COCOmetric (mAP)
Published: Medium, August 2021.
https://medium.com/@yrodriguezmd/object-detection-using-a-deep-neural-network-213ec8ac2da8
Focus: Object detection using faster rcnn, yolo5, retinanet, efficientdet; class label adjustment
Published: Medium, August 2021,
https://medium.com/@yrodriguezmd/different-models-for-object-detection-9c5cda7863c1
Finding: Fine_tune has a cumulative effect, LR gets smaller on subsequent iterations
Focus: Using IceVision voc dataset and parser, modelling with faster rcnn, yolov5, retinanet and efficientdet
Focus: Using plantsdoc tensorflow OD csv from roboflow, local computer download and upload, making a custom parser
Published: Medium, 2021, September
https://medium.com/@yrodriguezmd/the-custom-parser-a-key-to-a-good-data-harvest-10e24d0d8a71
Focus: Using git clone, with separate csv and images, making a custom parser
Mentioned in: https://medium.com/@yrodriguezmd/the-custom-parser-a-key-to-a-good-data-harvest-10e24d0d8a71
Focus: Revision of airctic/icedata/notebooks/dev/plantdoc.ipynb to update codes, tutorial
Publication: submitted for open-source PR, 2021 September
https://github.com/airctic/icedata/tree/master/notebooks/dev
Focus: Using plantdoc, uploaded via git clone, custom parser and modelling using faster rcnn, yolov5, retinanet and efficientdet
Publication: Medium, 2021 September
https://medium.com/@yrodriguezmd/modelling-for-leaf-disease-detection-e16554a14bee
Focus: Using a revised BCCD dataset (better annotation), upload via git clone, voc parser, no explicit class_map, modelling (final with yolov5), callbacks, save. (Fit_one_cycle not included-> placed in IV_in_the_works)
Publication: Medium, 2021 September
https://medium.com/@yrodriguezmd/a-close-look-on-modelling-blood-cells-4625e832311f
Focus: BCCD Dataset for VOCBBoxParsing and modelling. Based on IV BCCD new.
Publication: submitted for open-source PR, 2021 September as airctic/icedata/notebooks/dev/bccd_rev
https://github.com/airctic/icedata/pulls
Branch path: notebooks/dev/bccd_dev.ipynb
Focus: Global wheat dataset, custom parsing and modelling (with extension via fit_one_cycle).
IV_wheat_kaggle_works!! shows custom parsing using 'source' as class. Final model using Faster R-CNN.
IV_wheat_kaggle_yolov5 shows custom parsing with single class 'wheat'. Final model using YOLOv5, mAP 0.46 after 40 epochs (still with room to increase, but should be adequate for wheat head detection purposes).
Publication: Medium, September 2021.
https://medium.com/@yrodriguezmd/artificial-intelligence-can-help-feed-humans-9233f1c941f6
Focus: Simple custom Parser for a single-object type detection, based on wheat dataset.
Publication: submitted for open-source PR, 2021 September as airctic/icedata/notebooks/custom_parser
Focus: Using a pretrained Retinanet model for inference of bboxes and object_ids, creating a coco json file, refining the annotations in roboflow and parsing the resulting image and annotations.
Publication: pending