Stitcher

This project provides the implementation for Stitcher: Feedback-driven Data Provider for Object Detection. In this paper, we present Stitcher, a feedback-driven data provider, which aims to train object detectors in a balanced way. In Stitcher, images are resized into smaller components and then stitched into the same size to regular images. Stitched images contain inevitable smaller objects, which is exploited with the loss statistics as feedback to guide next-iteration update. Stitcher steadily improves performance by a large margin in all settings, especially for small objects, with nearly no additional computation in both training and testing stages.

Installation

• This project is based on maskrcnn-benchmark. Please check INSTALL.md for installation instructions.

Training and Inference

• Training and inference can be conducted with corresponding default scripts in maskrcnn-benchmark.

Trained Models

Faster R-CNN with FPN

Model	period	AP	AP small	Model
ResNet-50	1x	38.6	24.4	GoogleDrive
ResNet-50 (VOC)	1x	81.6	-	GoogleDrive
ResNet-101	1x	40.8	25.8	GoogleDrive
ResNet-50	2x	39.9	25.1	GoogleDrive
ResNet-101	2x	42.1	26.9	GoogleDrive
ResNext-101	1x	43.1	28.0	GoogleDrive
ResNet-101 + DCN	1x	43.3	27.1	GoogleDrive
ResNext-101 + DCN	1x	45.4	29.4	GoogleDrive

RetinaNet

Model	period	AP	AP small	Model
ResNet-50	1x	37.8	22.1	GoogleDrive
ResNet-101	1x	39.9	24.7	GoogleDrive
ResNet-50	2x	39.0	23.4	GoogleDrive
ResNet-101	2x	41.3	25.4	GoogleDrive

Mask R-CNN

Model	period	AP	AP small	Model
ResNet-50-FPN	1x	35.1	17.0	GoogleDrive
ResNet-101-FPN	1x	37.2	19.0	GoogleDrive

Longer periods

Model	period	AP	AP small	Model
Baseline	6x	35.6	19.8	GoogleDrive
Stitcher	6x	40.4	26.1	GoogleDrive

Selection paradigms (Faster R-CNN with FPN in 1x)

Model	Feedback	AP	AP small	Model
ResNet-50	Input	38.1	23.1	GoogleDrive
ResNet-50	Cls Loss	38.5	23.9	GoogleDrive
ResNet-50	Reg Loss	38.6	24.4	GoogleDrive
ResNet-50	Both Loss	38.5	23.7	GoogleDrive

Batch Dimension Stitcher (Faster R-CNN with FPN in 1x)

Model	k	AP	AP small	Model
ResNet-50	2	38.3	22.9	GoogleDrive
ResNet-50	3	38.5	22.9	GoogleDrive
ResNet-50	4	38.6	23.4	GoogleDrive
ResNet-50	5	38.7	23.7	GoogleDrive
ResNet-50	6	38.6	23.5	GoogleDrive
ResNet-50	7	38.4	23.6	GoogleDrive
ResNet-50	8	38.3	24.3	GoogleDrive

• These models are uploading.

Citation

Please cite Stitcher if it helps your research.

@misc{chen2019detnas,
    title={Stitcher: Feedback-driven Data Provider for Object Detection},
    author={Yukang Chen, Peizhen Zhang, Zeming Li, Yanwei Li, Xiangyu Zhang, Gaofeng Meng, Shiming Xiang, Jian Sun, Jiaya Jia},
    year={2020},
    booktitle = {arxiv},
}