In this repo the code belonging to my master thesis titled: "Active Learning for Object Detection With Localization Uncertainty from Sampling-Based Probabilistic Bounding Boxes" can be found. I also uploaded the pdf of my thesis here. As a very short summary: I researched the use of a localization uncertainty, obtained trough an ensemble of object detectors to select more informative images to be labeled. It shows promissing results on Pascal VOC 2007, but has not been used on other datasets. Please let me know your experiences if you use it on different datasets.

As a basis for my repository I used the excellent repository by Max de Groot and Ellis Brown PyTorch implementation of the SSD detector, retrieved on 19-02-2019. However, as I used the then newest stable version of PyTorch (1.0.1), I did change some of their code to be able to run it. Note that their repo is probably more suitable if you just want to use an SSD written in PyTorch and don't want to perform acive learning. Some parts of this readme are directly copy-pasted from Max de Groot and Ellis Brown their repo as my work is built upon their code anyways.

After finishing my thesis, in order to make it more useable for others, I cleaned the code a bit and wrote this readme. I hope this helps, however, bear in mind that the code is research code and should be viewed as such. Currently I'm traveling trough Central and South America. I know the code could still be improved but it should work if you follow the instructions below. Please post issues if you are serious about using it and don't understand certain parts. I'll see what I can do when I'm back.

• Getting Started
• Datasets
• Reference

A PyTorch implementation of Single Shot MultiBox Detector from the 2016 paper by Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang, and Alexander C. Berg. The official and original Caffe code can be found here.

• I supplied a list of the conda environment I used for my experiments in the requirements file for reproducability. The most important packages are probably: PyTorch, NumPy, SciPy, cv2 and hdbscan.
• Clone this repository.
• Then download the dataset by following the instructions below. Note that the Active Learning code has only been completely implemented for the Pascal VOC 2007 dataset.
• As the SSD uses a reduced VGG-16 backbone, download the fc-reduced VGG-16 PyTorch base network weights at: https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth and put them in a directory called 'weights'
• By default, we assume you have downloaded the file in the Active_Learning_Object_Detection/weights dir:

mkdir weights
cd weights
wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth

• Note that a GPU is highly recommended for training the SSD.

• As there are many parser argument options, I provide two sample run scripts in the run_scripts directory to get a headstart. I provided one for the six classes I used in my experiments and one for a single class of interest (thus background vs non-background). To use them, you also need to copy the appropriate imageset files to the imageset folder. You can find the imageset files in data/imageset_files and they need to be copied to `~/data/VOCdevkit/VOC2007/ImageSets/Main/.

• You are required to give a list of paths to the currently best networks. I provided a script create_initial_networks.py to generate these if you don't have any yet. The current settings of this script correspond to the sample run script with the single class. NOTE: A single saved network requires approximately 100MB of storage. Make sure you have enough diskspace before running the script.

• Finally, if you don't want to use one of the provided scripts, the entry point for active learning is the active_learning_main.py file.

For active learning the two most important folders are the active_learning_dir and active_learning_package. In the first the (intermediate) results of the runs (e.g. which images to label next) will be saved and in the second the code for the active learning can be found.

To make things easy, we provide bash scripts to handle the dataset (Pascal VOC) downloads and setup for you. We also provide simple dataset loaders that inherit torch.utils.data.Dataset, making them fully compatible with the torchvision.datasets API.

PASCAL VOC: Visual Object Classes

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory>

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>

• We are trying to provide PyTorch state_dicts (dict of weight tensors) of the latest SSD model definitions trained on different datasets.
• Currently, we provide the following PyTorch models:
  • SSD300 trained on VOC0712 (newest PyTorch weights)
    • https://s3.amazonaws.com/amdegroot-models/ssd300_mAP_77.43_v2.pth
  • SSD300 trained on VOC0712 (original Caffe weights)
    • https://s3.amazonaws.com/amdegroot-models/ssd_300_VOC0712.pth

Active learning part:

• Jasper Bakker

SSD, Dataloaders, etc. (check their excellent repo at PyTorch implementation of the SSD detector):

• Max deGroot
• Ellis Brown