GAIA-vision / GAIA-det

A Transfer Learning System in Object Detection

GAIA-det

More models and demos coming soon! Stay tuned.

Introduction

GAIA-det is an open source object detection toolbox that helps you with your customized AI solutions. It is built on top of gaiavision and mmdet. This repo includes an official re-implementation of our CVPR2021 paper:

• GAIA: A Transfer Learning System of Object Detection that Fits Your Needs.

It provides functionalities that help the customization of AI solutions.

• Design customized search space of any type with little efforts.
• Manage models in search space according to your rules.
• Integrate datasets of various sources.

Requirements

• Python 3.6+
• CUDA 10.0+
• 1.2.7 <= mmcv < 1.3.0
• 2.8.0 <= mmdet < 2.9.0
• Others (See requirements.txt)

Installation

• Install gaiavision.
• Install mmdet.
• Install gaiadet:

git clone https://github.com/GAIA-vision/GAIA-det.git && cd GAIA-det
pip install -r requirements.txt
pip install -e .

Prepare Supernet

Type	Backbone	Model	Data	Cloud Storage	Password
WEIGHTS	AR50to101	Faster	COCO+Obj365	BaiduCloud	tm5n
FLOPS_LUT	AR50to101	Faster	COCO+Obj365	BaiduCloud	ttwq
WEIGHTS	AR50to101	Faster	COCO+Obj365+OID500	Coming soon
FLOPS_LUT	AR50to101	Faster	COCO+Obj365+OID500	Coming soon

Benchmark

Finetuning(Upstream-COCO)

Backbone	Pretrain	Model	Depth	Width	Input Scale	Lr schd	FLOPS	box AP (paper)	box AP (repo)
ResNet50	ImageNet	Faster	3, 4, 6, 3	64, 64, 128, 256, 512	800	1x	139G	37.1	37.6
ResNet50	ImageNet	Faster	3, 4, 6, 3	64, 64, 128, 256, 512	800	4x	139G	None	40.3
45-50GF	GAIADET	Faster	2, 4, 5, 3	64, 64, 96, 192, 384	480	1x	49G	40.4	40.7
70-75GF	GAIADET	Faster	4, 6, 27, 4	48, 64, 128, 192, 512	480	1x	71G	42.6	43.1
85-90GF	GAIADET	Faster	3, 4, 21, 4	48, 64, 160, 192, 640	560	1x	90G	43.6	44.4
110-115GF	GAIADET	Faster	2, 4, 25, 4	64, 64, 160, 192, 640	640	1x	115G	44.5	44.8
135-140GF	GAIADET	Faster	4, 4, 15, 4	48, 48, 128, 192, 512	800	1x	139G	45.3	45.6

We compare our results with ResNet50 of 4x on COCO for fairness, because COCO data has been used for 3x during upstream training.

Compatibility with other methods

Backbone	Pretrain	Model	Depth	Width	Input Scale	Lr schd	Methods	box AP (paper)	box AP (repo)
ResNet50	ImageNet	Faster	3, 4, 6, 3	64, 64, 128, 256, 512	800	1x	N	37.1	37.6
ResNet50	ImageNet	Faster	3, 4, 6, 3	64, 64, 128, 256, 512	800	1x	Y	45.8	44.5
135-140GF	GAIADET	Faster	4, 4, 15, 4	48, 48, 128, 192, 512	800	1x	N	45.3	45.6
135-140GF	GAIADET	Faster	4, 4, 15, 4	48, 48, 128, 192, 512	800	1x	Y	49.1	48.5

Methods denote Deformable Convolution and Cascaded Head.

Finetuning(Downstream-BDD100k)

Backbone	Model	Depth	Width	Input Scale	Lr schd	FLOPS	box AP (paper)	box AP (repo)
ResNet50	Faster	3, 4, 6, 3	64, 64, 128, 256, 512	800	1x	139G	None	30.1
45-50GF	Faster	3, 4, 5, 2	48, 64, 96, 192, 384	480	1x	49G	None	27.4
70-75GF	Faster	4, 2, 15, 2	48, 48, 128, 192, 512	560	1x	71G	None	29.5
85-90GF	Faster	2, 2, 15, 3	64, 64, 128, 192, 384	640	1x	87G	None	32.1
135-140GF	Faster	4, 6, 23, 3	48, 80, 128, 192, 512	720	1x	139G	None	32.9

Finetuning(Downstream-UODB)

Dataset	KITTI	VOC	WiderFace	LISA	Kitchen	DOTA	DeepLesion	Comic	Clipart	Watercolor	Avg.
ResNet50(paper)	67.1	81.5	62.1	90.0	89.5	68.3	57.4	45.5	31.2	53.4	64.6
GAIA(paper)	75.6	87.4	62.7	92.1	90.1	70.8	62.1	61.1	72.2	69.7	74.4
ResNet50(repo)
GAIA(repo)

FLOPS of all models are around 139GFLOPS, and the metric used above is AP50.

Data Preparation

Please refer to DATA_PREPARATION.

Usage

Please refer to USAGE for generic use.

Citation

If you like our work and use the code or models for your research or project, please star our repo and cite our work as follows.

@InProceedings{Bu_2021_CVPR,
    author    = {Bu, Xingyuan* and Peng, Junran* and Yan, Junjie and Tan, Tieniu and Zhang, Zhaoxiang},
    title     = {GAIA: A Transfer Learning System of Object Detection That Fits Your Needs},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {274-283}
}