universenet

UniverseNet

Introduction

UniverseNets are state-of-the-art detectors for universal-scale object detection. Please refer to our paper for details. https://arxiv.org/abs/2103.14027

@inproceedings{USB_shinya_BMVC2022,
  title={{USB}: Universal-Scale Object Detection Benchmark},
  author={Shinya, Yosuke},
  booktitle={British Machine Vision Conference (BMVC)},
  year={2022}
}

Example for fine-tuning

For fine-tuning from a COCO pre-trained model, please see this example.

Results and Models

Main results

Method	Backbone	Lr schd	Mem (GB)	Inf time (fps)	box AP	Download
ATSS+SEPC	R-50	1x	-	25.0	42.1	model
UniverseNet	R2-50	1x	5.1	17.3	46.7	model
UniverseNet	R2-50	2x	5.1	17.3	48.9	model
UniverseNet+GFL	R2-50	1x	5.3	17.6	47.5	model
UniverseNet+GFL	R2-50	2x	5.3	17.6	49.4	model
UniverseNet+GFL	R2-101	2x	8.5	11.9	50.8	model
UniverseNet 20.08d	R2-50	1x	5.8	17.3	48.6	model
UniverseNet 20.08d	R2-101	20e	9.1	11.7	50.9	model
UniverseNet 20.08d	R2-101	2x	9.1	11.7	50.6	model
UniverseNet 20.08	R2-50	1x	5.5	24.9	47.5	model
UniverseNet 20.08	R2-50	2x	5.5	24.9	48.5	model

• In addition to ATSS+SEPC, UniverseNet uses Res2Net-v1b-50, DCN, and multi-scale training (480-960).
• The settings for normalization layers (including whether to use iBN of SEPC) depend on the config files.
• All models except for ATSS+SEPC were trained and evaluated using fp16 (mixed precision).
• The above UniverseNet (2x) model is a checkpoint at epoch 23. The AP of a checkpoint at epoch 24 is quite similar (48.9) but slightly worse.

Faster models

Method	Backbone	Test scale	Lr schd	Mem (GB)	Inf time (fps)	box AP	Download
UniverseNet 20.08s	R-50-C	(1333, 800)	2x	4.7	31.6	46.9	model
UniverseNet 20.08s	R-50-C	(512, 512)	2x	5.7	36.0	41.8	model
UniverseNet 20.08s	R-50-C	(224, 224)	2x	1.6	-	29.4	model

• 4 GPUs x 16 samples_per_gpu for small test scales ((512, 512) and (224, 224)). You will be able to use BatchNorm with norm_eval=False even on 1 GPU.

Test scale and test-dev AP

Method	Backbone	Max test scale	TTA	Inf time (fps)	box AP (val)	box AP (test-dev)
UniverseNet	R2-50	(1333, 800)	-	15.8	48.9	49.2
UniverseNet	R2-50	(1600, 960)	-	14.5	49.2	49.5
UniverseNet 20.08s	R-50-C	(1333, 800)	-	31.6	46.9	47.4
UniverseNet 20.08	R2-50	(1333, 800)	-	24.9	48.5	48.8
UniverseNet 20.08d	R2-101	(1333, 800)	-	11.7	50.9	51.3
UniverseNet 20.08d	R2-101	(2000, 1200)	5	-	53.1	53.8
UniverseNet 20.08d	R2-101	(3000, 1800)	13	-	53.5	54.1

• TTA: test-time augmentation including horizontal flip and multi-scale testing (numbers denote scales).

Misc.

Other hyperparameters and details for reproduction

Method	warmup_iters	lcconv_padding	GPUs x samples_per_gpu	box AP
UniverseNet	500	0	4x4 -> 8x2	48.9
UniverseNet	1000	1	4x4	48.9
UniverseNet	3665	0	4x4	48.8

• The checkpoints in release 20.06 were trained with a warmup_iters of 500. To make training more stable, the current config sets warmup_iters to 1000. The difference will not affect the final accuracy so much.
• In the official SEPC implementation, padding values in lconv and cconv (we call lcconv_padding) are set to 0. Setting lcconv_padding to 1 doesn't affect accuracy.
• To accelerate training for CVPR competitions, we used 8 GPUs for 9-24 epochs, after using 4 GPUs for 1-8 epochs.