Computer vision:

Object Detection to detect and count specific objects (shipping containers) from images and/or videos

This project is based on ChainerCV API and Single Shot MultiBox Detector algorithm. The dataset used for training is a mix of COCO dataset and manually labeled images (using yuyu21's tool). The reason for this is that my first attempt, using only my labeled images, was very effective in terms of True Positives, but was also generating lots of False Positives, so I had to enrich the datasets for some negative mining.

Labeling

Labeling is done throug the graphical interface (credit to MIT ) with the command:

python source/flask_app.py --image_dir datasets/train/ --label_names source/label_names_coco_container.yml --file_ext jpg

Training

I have been using a RedHat 7.4 server with a GPU (Tesla) for the training, and have gone through 500K+ iterations, with validation every 1K and decreased learning rate every 100K.

Training is done with the command:

python source/examples/ssd/train.py --train datasets/train/   --lr 0.0001 --step_size 100000  --val datasets/validation   --label source/label_names_coco_container.yml    --out models/ --gpu 0 --loaderjob 10   --iteration 1000000 --val_iteration 1000

Test on images

I created a Shiny app for the application of the model, which calls a python program to display side by side the original image and the result:

Test on videos

I've also downloaded a few clips of videos and labeled them. Since the model used is not very fast, it can't be used real time (detection takes a couple seconds per images).

So the script captures all the frames, then labels each of them, and reconstructs the video from the label frames:

python Video/label_video.py --video_file "path_to_video(s)_to_labelize"

Results