Awesome

TALISMAN: Targeted Active Learning for Object Detection with Rare Classes and Slices using Submodular Mutual Information

Official PyTorch implementation of ECCV 2022 paper "TALISMAN: Targeted Active Learning for Object Detection with Rare Classes and Slices using Submodular Mutual Information".

Suraj Kothawade, Saikat Ghosh, Sumit Shekhar, Yu Xiang, Rishabh Iyer

Overview

Deep neural networks based object detectors have shown great success in a variety of domains like autonomous vehicles, biomedical imaging, etc. It is known that their success depends on a large amount of data from the domain of interest. While deep models often perform well in terms of overall accuracy, they often struggle in performance on rare yet critical data slices. For example, data slices like "motorcycle at night" or "bicycle at night" are often rare but very critical slices for self-driving applications and false negatives on such rare slices could result in ill-fated failures and accidents. Active learning (AL) is a well-known paradigm to incrementally and adaptively build training datasets with a human in the loop. However, current AL based acquisition functions are not well-equipped to tackle real-world datasets with rare slices, since they are based on uncertainty scores or global descriptors of the image. We propose TALISMAN, a novel framework for Targeted Active Learning or object detectIon with rare slices using Submodular MutuAl iNformation. Our method uses the submodular mutual information functions instantiated using features of the region of interest (RoI) to efficiently target and acquire data points with rare slices. We evaluate our framework on the standard PASCAL VOC07+12 and BDD100K, a real-world self-driving dataset. We observe that TALISMAN outperforms other methods by in terms of average precision on rare slices, and in terms of mAP.

Prerequisites

Install PyTorch

Refer to pytorch official website https://pytorch.org/get-started/locally/

Install MMCV

pip install mmcv-full

Install Submodlib

pip install -i https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ submodlib

Clone Talisman repo from Github & Install dependencies

git clone https://github.com/surajkothawade/talisman.git # clone repository
cd talisman                                              # move inside talisman root directory
mkdir data                                               # create empty data directory
pip install -r requirements/build.txt                    # install build dependency
pip install -r requirements/runtime.txt                  # install runtime dependency
pip install -v -e .                                      # install talisman

Data Preparation

Prepare VOC Dataset

Download Pascal-VOC 07+12 dataset

wget http://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
wget http://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
wget http://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar

Unzip compressed dataset inside ‘talisman/data’ folder

tar -xvf '<path to VOCtrainval_06-Nov-2007.tar>' -C '<path to talisman/data/>'
tar -xvf '<path to VOCtrainval_11-May-2012.tar>' -C '<path to talisman/data/>'
tar -xvf '<path to VOCtest_06-Nov-2007.tar>'     -C '<path to talisman/data/>'

Prepare BDD Dataset

• Download 100K_images zip and Detection_2020 zip from https://bdd-data.berkeley.edu/
• Unzip bdd100k_images_100k.zip & bdd100k_det_20_labels_trainval.zip
• Convert det_20/labels/det_train.json & det_20/labels/det_val.json to coco format (Refer to https://doc.bdd100k.com/format.html#to-coco)
• Convert coco annotations to voc format by running below commands in order -

pip install imgann
from imgann import Convertor
Convertor.coco2voc('<path to images/100k/train>', '<path to labels/det_20/det_train_coco.json>', '<train_voc annotation output path>', False)
Convertor.coco2voc('<path to images/100k/val>', '<path to labels/det_20/det_val_coco.json>', '<val_voc annotation output path>', False)

• Create empty folder structure inside talisman/data as below and copy images, label annotations & .txt helper files in respective folders as shown below -

Set Experiment Parameters in scripts inside talisman/strategies

Set training parameters within the script as per requirement (line 39-46)

• set initialTraining = True if want to create first round model and labelled dataset
• if first round model and labelled dataset already present, keep initialTraining = False which is default

Set config_filename as per experiment (line 61)

• Work directory gets created as per the config file name. Hence, this should be unique for each experiment to avoid file and model overwrite. Example:

config_filename = 'faster_rcnn_r50_fpn_AL_voc0712.py'            # VOC - Rare experiment
config_filename = 'faster_rcnn_r50_fpn_AL_bdd100k_mc_rare.py'    # BDD - Rare Motorcycle
config_filename = 'faster_rcnn_r50_fpn_AL_bdd100k_ped_night.py'  # BDD - Pedestrian at night

Set split_cfg parameter (line 128) for rare class imbalance(VOC and BDD) & rare slice imbalance(BDD) as per below table

Set imbalanced_classes parameter (line 137) for rare class imbalance(VOC and BDD) & rare slice imbalance(BDD) as per below table –

Set rare slice parameters (line 140-142) only for BDD rare slice experiments

• set attr_property as either of timeofday (day/night slice), weather (rainy/clear) or scene (city streets/highway etc)
• set attr_value as per attr_property selected. Example:

Run Experiments

• For any experiment, run the first strategy with parameter initialTraining = True to create first round model and initial labelled seed set
• Once the First Round Training is complete and first round model Round_1.pth is available inside talisman/work_dirs/<config_filename>, other strategies can be run with parameter initialTraining = False.
• Different experiments and different strategies should run with different settings and with different argument list as described in the below table –