IFSOD-dataset

Dataset approched by A Benchmark and Frequency Compression Method for Infrared Few-Shot Object Detection

This dataset aims to provide an open evaluation scheme for small sample object detection tasks. It aims to enrich the training categories and sample pose changes. Contains over 4800 images, over 23000 instances, covering 8 scenes and 18 coarse-grained categories. We have established three random partitions for categories. Each partition contains base classes and new class images that comply with the few sample setting.

1. Dataset Overview

Provide detailed information about the dataset, including:

• Number of samples: 4800+ images
• Data format: images in JPG format and XML annotations

2. Dataset Samples

Below are some sample examples from the dataset to give users a better idea of its structure and content.

Sample

• Description: This sample shows [what the image/data represents]. The corresponding label is [label name or value].

3. Split

Class	Split I	Split II	Split III
Base Class
	Armored Car (217)	Kettle (122)	People (14860)
	Car (4167)	Pram (196)	Umbrella (217)
	Bike (716)	Goose (193)	Luggage (614)
	Dog (166)	Bus (510)	Bike (716)
	People (14860)	Dog (166)	Pram (196)
	Switching (241)	Car (4167)	Etricycle (262)
	Truck (153)	Duck (112)	Ebike (568)
	Tricycle (139)	Switching (241)	Car (4167)
	Goose (193)	People (14860)	Duck (112)
	Kettle (122)	Tricycle (139)	Kettle (122)
	Etricycle (262)	Ebike (568)	Truck (153)
	Umbrella (217)	Etricycle (262)	Armored Car (217)
	Guidepost (240)	Umbrella (217)	Guidepost (240)
Novel Class
	Duck (10)	Armored Car (10)	Bus (10)
	Ebike (10)	Truck (10)	Switching (10)
	Pram (10)	Bike (10)	Tricycle (10)
	Bus (10)	Luggage (10)	Dog (10)
	Luggage (10)	Guidepost (10)	Goose (10)

4. Dataset Comparison Table

The table below provides a comparison of different datasets, including their images, instances, resolution, and other attributes.

Dataset	Images	Instances	Resolution	Instance Density	Classes	Scenes	Meaningless Classes	Unresolved Classes
RGB-T234	233,928	116,660	(628,459)	0.500	145	8	60	43
M3FD	9,200	34,408	(1001,744)	3.74	6	13	0	0
LLVIP	15,485	41,579	(1028,1024)	2.685	1	7	0	0
IFSOD-dataset (Ours)	4,815	23,333	(662,489)	4.846	18	12	0	0

5. Performance Comparison Table

This table compares the performance of different methods across three novel splits. Results are given for various numbers of shots (1, 2, 3, 5, 10). The best results and the second-best results are highlighted in bold.

Method	Venue	Backbone	1	2	3	5	10	1	2	3	5	10	1	2	3	5	10
			Novel Split 1	-----	-----	-----	Novel Split 2	-----	-----	-----	Novel Split 3	-----	-----	-----
FSRW	ICCV2019	YOLOv2	8.82	13.55	16.70	23.91	27.21	15.76	15.30	22.77	30.19	29.24	10.20	18.73	22.70	26.67	25.43
Meta R-CNN	ICCV2019	FRCN-101	2.52	9.30	13.34	16.34	14.80	4.00	9.82	9.70	7.56	13.68	8.16	10.82	17.04	15.88	17.52
TFA w/cos	ICML2020	FRCN-101	5.70	10.32	17.44	21.80	26.12	0.70	7.74	8.86	9.94	16.90	5.54	3.28	5.48	5.76	11.10
MPSR	ECCV2020	FRCN-101	9.82	33.04	38.14	45.30	48.48	24.54	25.58	20.72	30.46	43.96	9.78	22.00	36.40	41.84	49.12
FsDetView	ECCV2020	FRCN-101	1.82	13.86	15.86	15.16	14.72	4.00	7.98	10.20	7.99	9.94	3.82	7.96	14.52	15.76	15.10
KFSOD	CVPR2021	FRCN-101	13.65	22.47	36.44	43.33	58.54	4.82	8.45	30.17	37.11	44.75	21.68	28.43	40.22	38.80	54.79
FSCE	CVPR2021	FRCN-101	15.20	21.43	42.20	50.94	55.98	2.66	8.38	34.69	41.55	46.65	24.01	32.15	45.82	50.41	58.30
CME	CVPR2021	FRCN-101	6.37	10.76	39.52	44.06	49.79	6.38	7.10	26.50	30.97	37.74	15.17	23.32	27.55	39.79	45.01
FADI	NeurIPS2021	FRCN-101	11.53	23.88	36.09	47.86	52.57	22.85	27.54	32.19	42.35	44.11	25.32	35.29	42.80	45.49	49.83
DeFRCN	ICCV2021	FRCN-101	12.69	20.59	42.14	44.00	46.16	6.37	10.19	31.80	45.13	37.82	23.53	25.46	36.83	43.38	46.12
FCT	CVPR2022	PVTv2	9.97	28.28	43.85	53.91	57.89	25.77	34.72	45.72	50.16	55.32	29.00	34.21	46.87	51.72	54.96

Dataset Download

Important: The dataset is available for download from the link below. Make sure to cite this dataset if you use it in your research.

Baidu yunpan