Bayesian Detector Combination

ECCV 2024 | Poster and Presentation | Proceeding |

Description

Acquiring fine-grained object detection annotations in unconstrained images is time-consuming, expensive, and prone to noise, especially in crowdsourcing scenarios. Most prior object detection methods assume accurate annotations; A few recent works have studied object detection with noisy crowdsourced annotations, with evaluation on distinct synthetic crowdsourced datasets of varying setups under artificial assumptions. To address these algorithmic limitations and evaluation inconsistency, we first propose a novel Bayesian Detector Combination (BDC) framework to more effectively train object detectors with noisy crowdsourced annotations, with the unique ability of automatically inferring the annotators' label qualities. Unlike previous approaches, BDC is model-agnostic, requires no prior knowledge of the annotators' skill level, and seamlessly integrates with existing object detection models. Due to the scarcity of real-world crowdsourced datasets, we introduce large synthetic datasets by simulating varying crowdsourcing scenarios. This allows consistent evaluation of different models at scale. Extensive experiments on both real and synthetic crowdsourced datasets show that BDC outperforms existing state-of-the-art methods, demonstrating its superiority in leveraging crowdsourced data for object detection.

(a)	(b)	(c)
(d)	(e)	(f)

Figure 1: Examples of ambiguous cases with noisy or incorrect annotations on (a - c) MS COCO, (d)(e) VinDr-CXR and (f) disaster response dataset.

Figure 2: Overall architecture of the proposed BDC. The process of updating the aggregator's parameter (blue arrows) and the object detector's parameters (red arrows) is repeated iteratively until convergence.

Setup Environment

Disclaimer: The codes have only been tested on Ubuntu 18.04, Python 3.8.18 and Pytorch 1.13.1 CUDA 11.7, but they should work on environments with similar major versions.

1. Install required libraries by running

   conda create -n bdc python=3.8
   conda activate bdc
   pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
   pip install -r requirements.txt

   Alternatively, install from conda yaml file with conda env create -f env.yaml
2. Install optional libraries for EVA model
   • Install Apex v22.12

   # install xformers for eva02 https://github.com/facebookresearch/xformers/tree/v0.0.16
   pip install xformers==0.0.16
   # compile detectron2
   cd models/eva
   python -m pip install -e .
   cd ../../
   pip install mmcv-full==1.7.1 -f https://download.openmmlab.com/mmcv/dist/cu117/torch1.13.0/index.html

3. Download pretrained model weights (optional)

   mkdir pretrained_weights
   cd pretrained_weights
   wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7_training.pt
   wget https://huggingface.co/Yuxin-CV/EVA-02/resolve/main/eva02/det/eva02_L_coco_det_sys_o365.pth

Benchmark Dataset

Download our synthesised crowdsourced annotations from here. Extract and place the annotations in data directory. Please download Pascal VOC, MSCOCO and VinDr-CXR and place the images in .data directory following this structure:

.
└── .data/
    ├── MSCOCO/
    │   ├── annotations
    │   └── images
    ├── VINCXR/
    │   ├── train
    │   └── test
    └── VOCdevkit/
        └── VOC2007/
            ├── Annotations
            └── JPEGImages

Preparing custom datasets

1. Convert your annotations to YOLO labelling format (one txt file per image).
   • Train annotations follows format: x1,y1,x2,y2,class_id,annotator_id
   • Test annotations (if available, else create an empty text file for each test image) follows format: x1,y1,x2,y2,class_id
2. Create a new yaml file and modify the settings in it following the examples in data folder.

How to run

Run the scripts with --help argument to see arguments descriptions:

• Specify the aggregation method with --crowd-aggregator argument, available aggregators are in config directory
• Specify the project name and experiment name with --project and --name arguments (experiment outputs are saved to outputs/<project>/<name>)
• Specify the data yaml file with --data argument

Examples below train YoloV7, Faster R-CNN and EVA model on our simplest synthetic dataset: VOC-FULL using our proposed BDC aggregation method

python train_yolov7.py --project crowd-annotations-voc --plots --batch-size 31 --weights ./pretrained_weights/yolov7_training.pt --crowd-aggregator config/bdc.yaml --name ann10_full-yolov7_pretrained-bdc --data data/voc_2007_ann10_full.yaml
python train_faster_rcnn.py --project crowd-annotations-voc --plots --batch-size 8 --crowd-aggregator config/bdc.yaml --name ann10_full-frcnn_pretrained-bdc --data data/voc_2007_ann10_full.yaml
python train_eva.py --project crowd-annotations-voc --plots --crowd-aggregator config/bdc.yaml --name ann10_full-eva_pretrained-bdc --data data/voc_2007_ann10_full.yaml

Examples below train YoloV7, Faster R-CNN and EVA model on our synthetic dataset COCO-MIX using our proposed BDC aggregation method

python train_yolov7.py --cfg ./models/yolov7/config/yolov7_coco.yaml --epochs 20  --save-after 5 --save-interval 5 --project crowd-annotations-coco --plots --batch-size 31 --weights ./pretrained_weights/yolov7_training.pt --crowd-aggregator config/bdc.yaml --name ann1000_mix-yolov7_pretrained-bdc --data data/coco_ann1000_mix_disjoint.yaml
python train_faster_rcnn.py --project crowd-annotations-coco --plots --epochs 10 --save-after 5 --save-interval 1 --batch-size 8 --crowd-aggregator config/bdc.yaml --name ann1000_mix-frcnn_pretrained-bdc --data data/coco_ann1000_mix_disjoint.yaml
python train_eva.py --project crowd-annotations-coco --plots --epochs 10 --crowd-aggregator config/bdc.yaml --name ann1000_mix-eva_pretrained-bdc --data data/coco_ann1000_mix_disjoint.yaml

All experiments are logged to Tensorboard and WandB automatically. They can be disabled in the training codes.

Crowdsourced Annotations Synthesisation

The code to synthesise the synthetic datasets are implemented in Jupyter notebooks in notebook directory. Please refer to the notebooks synthetic_*.ipynb for more details. Synthesisation code (with docstrings) can also be found in utils.crowd.synthetic_data.py.

General steps to synthesise annotations are as follows:

1. Prepare your custom dataset following instructions above
2. Run train_classification.py to train the classification model and test_classification.py to obtain the confusion matrix for synthesising class labels
3. Run train_rpn.py to train the RPN model for synthesising bounding boxes
4. Follow the steps in the notebooks to synthesise annotations and modify according to your desired synthetic setting (changing the number of annotators, classification and RPN models, etc.)

Citation

If you find this work useful for your research, please cite our work as

@inproceedings{bdc2024tan,
   title     = {Bayesian Detector Combination for Object Detection with Crowdsourced Annotations},
   author    = {Tan, Zhi Qin and Isupova, Olga and Carneiro, Gustavo and Zhu, Xiatian and Li, Yunpeng},
   booktitle = {Proc. Eur. Conf. Comput. Vis.},
   pages     = {329--346},
   year      = {2024},
   address   = {Milan, Italy},
}

Feedback

Suggestions and opinions on this work (both positive and negative) are greatly welcomed. Please contact the authors by sending an email to zhiqin1998 at hotmail.com.

Acknowledgement

This work is developed based on the codebase of YOLOv7 and EVA. We thank the authors for releasing their source code and models.