Repository files navigation

Official PyTroch implementation of PPDet based on open-mmlab's mmdetection.

Reducing Label Noise in Anchor-Free Object Detection,
Nermin Samet, Samet Hicsonmez,Emre Akbas,
BMVC 2020. (arXiv pre-print)

Current anchor-free detectors label all features within a ground-truth box as positive. However features within a ground-truth box may come from (i) the background, (ii) occluders or (iii) non-discriminatory parts of the object. PPDet avoids treating such features as positive. For example in the image below, the colored dots show the locations whose predictions are pooled to generate the final detection shown in the green bounding box. The color denotes the contribution weight. Highest contributions are coming from the objects and not occluders or background areas.

Current anchor-free object detectors label all the features that spatially fall inside a predefined central region of a ground-truth box as positive. This approach causes label noise during training, since some of these positively labeled features may be on the background or an occluder object, or they are simply not discriminative features. In this paper, we propose a new labeling strategy aimed to reduce the label noise in anchor-free detectors. We sum-pool predictions stemming from individual features into a single prediction. This allows the model to reduce the contributions of non-discriminatory features during training. We develop a new one-stage, anchor-free object detector, PPDet, to employ this labeling strategy during training and a similar prediction pooling method during inference. On the COCO dataset, PPDet achieves the best performance among anchor- free top-down detectors and performs on-par with the other state-of-the-art methods. It also outperforms all state-of-the-art methods in the detection of small objects (AP_s 31.4).

• PPDet (Prediction Pooling Detector) is a new relaxed labelling strategy for anchor-free object detection.
• To reduce the contribution of non-discriminatory features during training, PPDetsum-pool predictions stemming from individual features into a single prediction.
• PPDet is uses a novel prediction pooling strategy in training and inference.
• PPDet is state-of-the-art method in the detection of small objects with AP_s 31.4.
• Our best model achieves46.3 AP on COCO test-dev.

• For multi scale testing we used scales of (800, 480), (1067, 640), (1333, 800), (1600, 960), (1867, 1120) and (2133, 1280).

PPDet is implemented on top ofmmdetection. Therefore the installation is the same as original mmdetection.

You could checkINSTALL.md for installation instructions.

The PPDet configs could be found inconfigs/ppdet.

# single-gpu testingpython tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] --eval bbox [--show]# multi-gpu testing./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [--out ${RESULT_FILE}] --eval bbox

# single-gpu trainingpython tools/train.py ${CONFIG_FILE}# multi-gpu training./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} [optional arguments]

This work was supported by the AWS Cloud Credits for Research program and by the Scientific and Technological Research Council of Turkey (TÜBİTAK) through the project titled "Object Detection in Videos with Deep Neural Networks" (grant number 117E054). The numerical calculations reported in this paper were partially performed at TÜBİTAK ULAKBİM, High Performance and Grid Computing Center (TRUBA resources).

PPDet is released under the Apache License (refer to theLICENSE file for details). We developed PPDet on top of open-mmlab'smmdetection. Please refer to the License ofmmdetection for more detail.

If you find PPDet useful for your research, please cite our paper as follows.

N. Samet, S. Hicsonmez, E. Akbas, "Reducing Label Noise in Anchor-Free Object Detection",In British Machine Vision Conference (BMVC), 2020.

BibTeX entry:

@inproceedings{PPDet,  author = {Nermin Samet and Samet Hicsonmez and Emre Akbas},  title = {Reducing {L}abel {N}oise in {A}nchor-{F}ree {O}bject {D}etection},  booktitle = {British Machine Vision Conference (BMVC)},  year = {2020},}