Skip to content

Latest commit

 

History

History
95 lines (80 loc) · 18.8 KB

MODEL_ZOO.md

File metadata and controls

95 lines (80 loc) · 18.8 KB
Raw

R3Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object

Performance（deprecated）

Due to the improvement of the code, the performance of this repo is gradually improving, so the experimental results in this file are for reference only.

Baseline

Model Backbone Training data Val data mAP Model Link Anchor Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
RetinaNet (baseline) ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 53.17 - H smooth L1 90 1x No 8X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v3.py
RetinaNet (baseline) ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 63.18 model H smooth L1 90 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v4.py
RetinaNet (baseline) ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 62.79 - H smooth L1 90 2x No 8X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v8.py
RetinaNet (baseline) ResNet101_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 64.73 - H smooth L1 90 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res101_dota_v9.py
RetinaNet (baseline) ResNet152_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 66.97 - H smooth L1 90 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res152_dota_v12.py
RetinaNet ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 65.11 - H smooth L1 + atan(theta) 90 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v16.py
RetinaNet ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 64.10 - H smooth L1 180 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v15.py
RetinaNet (baseline) ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 62.76 model R smooth L1 90 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v1.py
RetinaNet (baseline) ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 62.25 - R smooth L1 90 2x No 8X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v10.py
RetinaNet ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 68.65 - R iou-smooth L1 [-ln(x)] 90 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v5.py

R3Det

Model Backbone Training data Val data mAP Model Link Anchor Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
R3Det ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 66.31 - H + R smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v1.py
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 67.29 (67.66) - H + R smooth L1 90 2x No 2X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v2.py
R3Det ResNet101_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 71.69 - H + R smooth L1 90 3x Yes 8X GeForce RTX 2080 Ti 1 -
R3Det* ResNet101_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 73.47 - H + R iou-smooth L1 [1-exp(1-x)] 90 3x Yes 4X GeForce RTX 2080 Ti 1 cfgs_res101_dota_r3det_v19.py
R3Det ResNet152_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 72.81 - H + R smooth L1 90 4x Yes 8X GeForce RTX 2080 Ti 1 -
R3Det* ResNet152_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 73.74 - H + R smooth L1 90 4x Yes 8X GeForce RTX 2080 Ti 1 -
R3Det* ResNet152_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 73.91 - H + R iou-smooth L1 [1-exp(1-x)] 90 3x Yes 4X GeForce RTX 2080 Ti 1 cfgs_res152_dota_r3det_v25.py

Anchor Free

Model Backbone Training data Val data mAP Model Link Anchor Anchor Scale Anchor Ratio Positive Threshold Negative Threshold Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 67.66 - H + R [2 ** 0, 2 ** (1.0 / 3.0), 2 ** (2.0 / 3.0)] [1, 1 / 2, 2., 1 / 3., 3., 5., 1 / 5.] [0.5, 0.6, 0.7] [0.4, 0.5, 0.6] smooth L1 90 2x No 2X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v2.py
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 65.89 - H + R [1.] [1.] [0.5, 0.6, 0.7] [0.4, 0.5, 0.6] smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v9.py
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 67.11 - H + R [2 ** 0, 2 ** (1.0 / 3.0), 2 ** (2.0 / 3.0)] [1.] [0.35, 0.5, 0.6] [0.25, 0.4, 0.5] smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v5.py
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 67.32 - H + R [1.] [1.] [0.35, 0.5, 0.6] [0.25, 0.4, 0.5] smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v7.py
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 69.50 - H + R [1.] [1.] [0.35, 0.5, 0.6] [0.25, 0.4, 0.5] iou-smooth L1 [1-exp(1-x)] 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v12.py

R3Det++

Model Backbone Training data Val data mAP Model Link InLD Anchor Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
R3Det++ ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 69.07 - {4,4,3,2,2} H + R smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_plusplus_v2.py
R3Det++ ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 69.81 - {1,1,1,1,1} H + R smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_plusplus_v3.py
R3Det++ ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 70.08 - {1,1,1,1,1} H + R iou-smooth L1 [1-exp(1-x)] 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_plusplus_v9.py
R3Det++ ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 68.12 - {1,1,1,1,1} + binary H + R smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_plusplus_v7.py
R3Det++ ResNet101_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 69.19 - {1,1,1,1,1} H + R smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res101_dota_r3det_plusplus_v4.py
R3Det++ ResNet101_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 72.98 - {1,1,1,1,1} H + R smooth L1 90 3x Yes 4X GeForce RTX 2080 Ti 1 cfgs_res101_dota_r3det_plusplus_v5.py
R3Det++ ResNet152_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 74.41 - {4,4,3,2,2} H + R smooth L1 90 4x Yes 8X GeForce RTX 2080 Ti 1 -
R3Det++ ResNet152_v1d MS DOTA1.0 trainval DOTA1.0 test 76.56 - {4,4,3,2,2} H + R + more smooth L1 90 6x Yes 4X GeForce RTX 2080 Ti 1 cfgs_res152_dota_r3det_plusplus_v1.py

IoU-Smooth L1 Loss

Model Backbone Training data Val data mAP Model Link Anchor Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
RetinaNet ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 62.25 - R smooth L1 90 2x No 8X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v10.py
RetinaNet ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 68.65 - R iou-smooth L1 [-ln(x)] 90 1x No 1X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v5.py
Model Backbone Training data Val data mAP Model Link Anchor Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
RetinaNet ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 62.79 - H smooth L1 90 2x No 8X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v8.py
RetinaNet ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 65.34 - H iou-smooth L1 [1-exp(1-x)] 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_v20.py
Model Backbone Training data Val data mAP Model Link Anchor Anchor Scale Anchor Ratio Positive Threshold Negative Threshold Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 67.32 - H + R (1.) (1.) (0.35, 0.5, 0.6) (0.25, 0.4, 0.5) smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v7.py
R3Det* ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 69.50 - H + R (1.) (1.) (0.35, 0.5, 0.6) (0.25, 0.4, 0.5) iou-smooth L1 [1-exp(1-x)] 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_v12.py
Model Backbone Training data Val data mAP Model Link InLD Anchor Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
R3Det++ ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 69.81 - {1,1,1,1,1} H + R smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_plusplus_v3.py
R3Det++ ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 70.08 - {1,1,1,1,1} H + R iou-smooth L1 [1-exp(1-x)] 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_plusplus_v9.py

EfficientNet

Model Backbone Training data Val data Anchor Anchor Scale Anchor Ratio Positive Threshold Negative Threshold Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU
R3Det* EfficientNet 600->800 DOTA1.0 trainval DOTA1.0 test H + R (1.) (1.) (0.35, 0.5, 0.6) (0.25, 0.4, 0.5) iou-smooth L1 [1-exp(1-x)] 90 2x No 2X GeForce RTX 2080 Ti 1
B0 B1 B2 B3 B4 B5 B6 B7 B8 L2-475 L2
Baseline preprocessing 61.84 (cfgs) 62.61 (cfgs) 63.00 (cfgs) 64.12 (cfgs) 64.72 (cfgs) 65.23 (cfgs)
AutoAugment (AA)
RandAugment (RA)
AdvProp + AA
NoisyStudent + RA 62.02 (cfgs) 67.44 (cfgs)

CSL

Model Backbone Training data Val data mAP Model Link InLD Anchor Label Mode Raduius/Sigma Reg. Loss Angle Range lr schd Data Augmentation GPU Image/GPU Configs
R3Det ResNet50_v1d 600->800 DOTA1.0 trainval DOTA1.0 test 64.88 - - H + R Gaussian 6 smooth L1 90 2x No 4X GeForce RTX 2080 Ti 1 cfgs_res50_dota_r3det_csl_v2.py

R3Det*: R3Det with two refinement stages
Some model results are slightly higher than in the paper due to retraining.