The Box Size Confidence Bias Harms Your Object Detector - Code

Disclaimer: This repository is for research purposes only. It is designed to maintain reproducibility of the experiments described in "The Box Size Confidence Bias Harms Your Object Detector".

Setup

Download Annotations

Download COCO2017 annotations for train, val, and test-dev from here and move them into the folder structure like this (alternatively change the config in config/all/paths/annotations/coco_2017.yaml to your local folder structure):

 .
 └── data
   └── coco
      └── annotations
        ├── instances_train2017.json
        ├── instances_val2017.json
        └── image_info_test-dev2017.json

Generate Detections (or Download them here)

Generate detections on the train, val, and test-dev COCO2017 set, save them in the COCO file format as JSON files. Move detections to data/detections/MODEL_NAME, see config/all/detections/default_all.yaml for all the used detectors and to add other detectors.
The official implementations for the used detectors are:

• CenterNet(HG) [Link]
• Faster-RCNN, YOLOv3-320, DETR, RetinaNet, CenterNet (RN-18), etc. [Link]
• SSD [Link]
• CenterNet2 [Link]
• YOLOv5x [Link]
• EfficientDet D0-7 [Link]

Examples

CenterNet (Hourglass)

To generate the Detections for CenterNet with Hourglass backbone first follow the installation instructions. Then download ctdet_coco_hg.pth to /models from the official source Then generate the detections from the /src folder:

# On val
python3 test.py ctdet --arch hourglass --exp_id Centernet_HG_val --dataset coco --load_model ../models/ctdet_coco_hg.pth 
# On test-dev
python3 test.py ctdet --arch hourglass --exp_id Centernet_HG_test-dev --dataset coco --load_model ../models/ctdet_coco_hg.pth --trainval
# On train
sed '56s/.*/  split = "train"/' test.py > test_train.py
python3 test_train.py ctdet --arch hourglass --exp_id Centernet_HG_train --dataset coco --load_model ../models/ctdet_coco_hg.pth

The scaling for TTA is set via the "--test_scales LIST_SCALES" flag. So to generate only the 0.5x-scales: --test_scales 0.5

RetinaNet with MMDetection

To generate the de detection files using mmdet, first follow the installation instructions. Then download specific model weights, in this example retinanet_x101_64x4d_fpn_2x_coco_20200131-bca068ab.pth to PATH_TO_DOWNLOADED_WEIGHTS and execute the following commands:

python3 tools/test.py configs/retinanet/retinanet_x101_64x4d_fpn_2x_coco.py PATH_TO_DOWNLOADED_WEIGHTS/retinanet_x101_64x4d_fpn_2x_coco_20200131-bca068ab.pth  --eval bbox --eval-options jsonfile_prefix='PATH_TO_THIS_REPO/detections/retinanet_x101_64x4d_fpn_2x/train2017' --cfg-options data.test.img_prefix='PATH_TO_COCO_IMGS/train2017' data.test.ann_file='PATH_TO_COCO_ANNS/annotations/instances_train2017.json'
python3 tools/test.py configs/retinanet/retinanet_x101_64x4d_fpn_2x_coco.py PATH_TO_DOWNLOADED_WEIGHTS/retinanet_x101_64x4d_fpn_2x_coco_20200131-bca068ab.pth  --eval bbox --eval-options jsonfile_prefix='PATH_TO_THIS_REPO/detections/retinanet_x101_64x4d_fpn_2x/val2017' --cfg-options data.test.img_prefix='PATH_TO_COCO_IMGS/val2017' data.test.ann_file='PATH_TO_COCO_ANNS/annotations/instances_val2017.json'
python3 tools/test.py configs/retinanet/retinanet_x101_64x4d_fpn_2x_coco.py PATH_TO_DOWNLOADED_WEIGHTS/retinanet_x101_64x4d_fpn_2x_coco_20200131-bca068ab.pth  --eval bbox --eval-options jsonfile_prefix='PATH_TO_THIS_REPO/detections/retinanet_x101_64x4d_fpn_2x/test-dev2017' --cfg-options data.test.img_prefix='PATH_TO_COCO_IMGS/test2017' data.test.ann_file='PATH_TO_COCO_ANNS/annotations/image_info_test-dev2017.json'

Install Dependencies

pip3 install -r requirements.txt

Optional Dependencies

# Faster coco evaluation (used if available)
pip3 install fast_coco_eval
# Parallel multi-runs, if enough RAM is available (add "hydra/launcher=joblib" to every command with -m flag)
pip install hydra-joblib-launcher

Experiments

Most of the experiments are performed using the CenterNet(HG) detections to change the detector add detections=OTHER_DETECTOR, with the location of OTHER_DETECTORs detections specified in config/all/detections/default_all.yaml. The results of each experiment are saved to outputs/EXPERIMENT/DATE and multirun/EXPERIMENT/DATE in the case of a multirun (-m flag).

Figure 2: Calibration curve of histogram binning and modified version

# original histogram binning calibration curve
python3 create_plots.py -cn plot_org_hist_bin
# modified histogram binning calibration curve:
python3 create_plots.py -cn plot_mod_hist_bin

Table 1: Ablation of histogram binning modifications

python3 calibrate.py -cn ablate_modified_hist 

Table 2: Ablation of optimization metrics of calibration on validation split

python3 calibrate.py -cn ablate_metrics  "seed=range(4,14)" -m

Figure 3: Bounding box size bias on train and val data detections

Plot of calibration curve:

# on validation data
python3 create_plots.py -cn plot_miscal name="plot_miscal_val" split="val"
# on train data:
python3 create_plots.py -cn plot_miscal name="plot_miscal_train" split="train" calib.conf_bins=20

Table 3: Ablation of optimization metrics of calibration on training data

python3 calibrate.py -cn explore_train

Table 4: Effect of individual calibration on TTA

1. Generate detections (on train and val split) for each scale-factor individually (CenterNet_HG_TTA_050, CenterNet_HG_TTA_075, CenterNet_HG_TTA_100, CenterNet_HG_TTA_125, CenterNet_HG_TTA_150) and for complete TTA (CenterNet_HG_TTA_ens)

2. Generate individually calibrated detections..

   python3 calibrate.py -cn calibrate_train name="calibrate_train_tta" detector="CenterNet_HG_TTA_050","CenterNet_HG_TTA_075","CenterNet_HG_TTA_100","CenterNet_HG_TTA_125","CenterNet_HG_TTA_150","CenterNet_HG_TTA_ens" -m

3. Copy calibrated detections from multirun/calibrate_train_tta/DATE/MODEL_NAME/quantile_spline_ontrain_opt_tradeoff_full/val/MODEL_NAME.json to data/calibrated/MODEL_NAME/val/results.json for MODEL_NAME in (CenterNet_HG_TTA_050, CenterNet_HG_TTA_075, CenterNet_HG_TTA_100, CenterNet_HG_TTA_125, CenterNet_HG_TTA_150).

4. Generate TTA of calibrated detections

   python3 enseble.py -cn enseble

Figure 4: Ablation of IoU threshold

python3 calibrate.py -cn calibrate_train name="ablate_iou" "iou_threshold=range(0.5,0.96,0.05)" -m

Table 5: Calibration method on different model

python3 calibrate.py -cn calibrate_train name="calibrate_all_models" detector=LIST_ALL_MODELS -m

The test-dev predictions are found in multirun/calibrate_all_models/DATE/MODEL_NAME/quantile_spline_ontrain_opt_tradeoff_full/test/MODEL_NAME.json and can be evaluated using the official evaluation sever.

Supplementary Material

A.Figure 5 & 6: Performance Change for Extended Optimization Metrics

python3 calibrate.py -cn ablate_metrics_extended  "seed=range(4,14)" -m

A.Table 6: Influence of parameter search spaces on performance gain

# Results for B0, C0
python3 calibrate.py -cn calibrate_train
# Results for B0, C1
python3 calibrate.py -cn calibrate_train_larger_cbins
# Results for B0 union B1, C0
python3 calibrate.py -cn calibrate_train_larger_bbins
# Results for B0 union B1, C0 union C1
python3 calibrate.py -cn calibrate_train_larger_cbbins

A.Table 7: Influence of calibration method on different sized versions of EfficientDet

python3 calibrate.py -cn calibrate_train name="influence_modelsize" detector="Efficientdet_D0","Efficientdet_D1","Efficientdet_D2","Efficientdet_D3","Efficientdet_D4","Efficientdet_D5","Efficientdet_D6","Efficientdet_D7" -m

Citation

@article{gilg2021box,
      title={The Box Size Confidence Bias Harms Your Object Detector}, 
      author={Johannes Gilg and Torben Teepe and Fabian Herzog and Gerhard Rigoll},
      year={2021},
      eprint={2112.01901},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}