laurentberder / container_detection

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

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

container_detection's Issues

Dataset

Hi sweetie,may I have your labeled dataset,please.

dataset

Hi, thanks for this repo! I'd like to experiment myself, would it be possible to get hold of your labelled dataset?

Can you upload the trained model?

Hey,

I loved your project. Is there a trained model that you could share with me? Thanks.

Container ID Detection and Recognition

Hallo Laurent,

first of all, thanks a lot for sharing your work. I find it very useful! I am working on a follow-up task and would like to use your dataset for it. The need is to read the IDs of containers, but very often they are not readable due to small image reolution in your dataset. Do you have any idea where I can get the image with the original resolution? That would be great, because then I could use your dataset for this, too.

Best regards,

André

Annotation video is not smooth(frames are too fast in anno video)

I use the voc0712 pretrained Chainer model and label is voc bbox label names and didn't use any custom model . Unable to attach the avi format file , So, I attached the original and annotation video here [https://github.com/vimal4353/chainer-video-obj-detection] . Please refer the video & let me know why the annotation video comes like this . I try with different fps (25,20,10). But didn't get the smoother video.