RESULTS.md

• 00.classification_training results
  • ResNetCifar training from scratch on CIFAR100
  • DarkNet training from scratch on ImageNet1K(ILSVRC2012)
  • ResNet training from scratch on ImageNet1K(ILSVRC2012)
  • ResNet finetune from ImageNet21k pretrain weight on ImageNet1K(ILSVRC2012)
  • Convformer finetune from offical pretrain weight on ImageNet1K(ILSVRC2012)
  • VAN finetune from offical pretrain weight on ImageNet1K(ILSVRC2012)
  • ViT finetune from self-trained MAE pretrain weight(400epoch) on ImageNet1K(ILSVRC2012)
  • ViT finetune from offical MAE pretrain weight(800 epoch) on ImageNet1K(ILSVRC2012)
  • ResNet train from pytorch official weight on ImageNet21K(Winter 2021 release)
• 01.distillation_training results
  • ResNet distill from pretrain weight on ImageNet1K(ILSVRC2012)
• 02.masked_image_modeling_training results
  • ViT MAE pretrain on ImageNet1K(ILSVRC2012)
• 03.detection_training results
  • All detection models training from scratch on COCO2017
  • All detection models finetune from objects365 pretrain weight on COCO2017
  • All detection models training from scratch on Objects365(v2,2020)
  • All detection models training from scratch on VOC2007&VOC2012
  • All detection models finetune from objects365 pretrain weight on VOC2007&VOC2012
• 04.semantic_segmentation_training results
• 05.instance_segmentation_training results
• 06.salient_object_detection_training results
• 07.human_matting_training results
• 08.ocr_text_detection_training results
• 09.ocr_text_recognition_training results
• 10.face_detection_training results
• 11.face_parsing_training results
• 12.human_parsing_training results
• 13.interactive_segmentation_training results
• 14.video_interactive_segmentation_training results
• 16.universal_segmentation_training
  • universal_segmentation semantic_segmentation_training results
  • universal_segmentation instance_segmentation_training results
  • universal_segmentation salient_object_detection_training results
  • universal_matting human_matting_training results
  • universal_matting human_instance_matting_training results
  • universal_segmentation face_parsing_training results
  • universal_segmentation human_parsing_training results

00.classification_training results

DarkNet

Paper:https://arxiv.org/abs/1804.02767?e05802c1_page=1

ResNet

Paper:https://arxiv.org/abs/1512.03385

Convformer

Paper:https://arxiv.org/abs/2210.13452

VAN

Paper:https://arxiv.org/abs/2202.09741

ViT

Paper:https://arxiv.org/abs/2010.11929

ResNetCifar training from scratch on CIFAR100

ResNetCifar is different from ResNet in the first few layers.

Model	input size	batch	epochs	Top-1
ResNet18Cifar	32x32	128	200	76.990
ResNet34Cifar	32x32	128	200	77.710
ResNet50Cifar	32x32	128	200	77.300
ResNet101Cifar	32x32	128	200	77.450
ResNet152Cifar	32x32	128	200	77.950

You can find more model training details in 00.classification_training/cifar100/.

DarkNet training from scratch on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Top-1
DarkNetTiny	256x256	256	100	58.074
DarkNet19	256x256	256	100	74.040
DarkNet53	256x256	256	100	76.366

You can find more model training details in 00.classification_training/imagenet/.

ResNet training from scratch on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Top-1
ResNet18	224x224	256	100	70.520
ResNet34	224x224	256	100	73.796
ResNet50	224x224	256	100	76.242
ResNet101	224x224	256	100	77.436
ResNet152	224x224	256	100	77.834

You can find more model training details in 00.classification_training/imagenet/.

ResNet finetune from ImageNet21k pretrain weight on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Top-1
ResNet50	224x224	2048	300	80.110
ResNet101	224x224	1024	300	81.586
ResNet152	224x224	1024	300	81.712

You can find more model training details in 00.classification_training/imagenet/.

Convformer finetune from offical pretrain weight on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Top-1
convformer-s18	224x224	2048	300	81.914
convformer-s36	224x224	2048	300	83.210
convformer-m36	224x224	1024	300	83.980
convformer-b36	224x224	1024	300	84.424

You can find more model training details in 00.classification_training/imagenet/.

VAN finetune from offical pretrain weight on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Top-1
van-b0	224x224	2048	300	75.216
van-b1	224x224	2048	300	80.608
van-b2	224x224	1024	300	82.540
van-b3	224x224	1024	300	83.240

You can find more model training details in 00.classification_training/imagenet/.

ViT finetune from self-trained MAE pretrain weight(400epoch) on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Top-1
ViT-Base-Patch16	224x224	256	100	82.794
ViT-Large-Patch16	224x224	128	50	84.842
ViT-Huge-Patch14	224x224	128	50	85.816

You can find more model training details in 00.classification_training/imagenet/.

ViT finetune from offical MAE pretrain weight(800 epoch) on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Top-1
ViT-Base-Patch16	224x224	256	100	83.152
ViT-Large-Patch16	224x224	128	50	85.870
ViT-Huge-Patch14	224x224	128	50	86.608

You can find more model training details in 00.classification_training/imagenet/.

ResNet train from pytorch official weight on ImageNet21K(Winter 2021 release)

Model	input size	batch	epochs	Semantic Softmax Acc
ResNet50	224x224	2048	80	75.354
ResNet101	224x224	2048	80	76.842
ResNet152	224x224	1024	80	77.342

You can find more model training details in 00.classification_training/imagenet21k/.

01.distillation_training results

DML loss

Paper:https://arxiv.org/abs/1706.00384

KD loss

Paper:https://arxiv.org/abs/1503.02531

ResNet distill from pretrain weight on ImageNet1K(ILSVRC2012)

Teacher Model	Student Model	method	Freeze Teacher	input size	batch	epochs	Teacher Top-1	Student Top-1
ResNet152	ResNet50	CE+DML	False	224x224	256	100	79.370	78.086
ResNet152	ResNet50	CE+DML+Vit Aug	False	224x224	1024	300	82.722	80.830
ResNet152	ResNet50	CE+KD	True	224x224	256	100	77.836	77.578
ResNet152	ResNet50	CE+KD+Vit Aug	True	224x224	2048	300	81.712	80.672

You can find more model training details in 01.distillation_training/imagenet/.

02.masked_image_modeling_training results

MAE

Paper:https://arxiv.org/abs/2111.06377

ViT MAE pretrain on ImageNet1K(ILSVRC2012)

Model	input size	batch	epochs	Loss
ViT-Base-Patch16	224x224	1024	400	0.3876
ViT-Large-Patch16	224x224	1024	400	0.3784
ViT-Huge-Patch14	224x224	1024	400	0.3502

You can find more model training details in 02.masked_image_modeling_training/imagenet/.

03.detection_training results

RetinaNet

Paper:https://arxiv.org/abs/1708.02002

FCOS

Paper:https://arxiv.org/abs/1904.01355

DETR

Paper:https://arxiv.org/abs/2005.12872

All detection models training from scratch on COCO2017

Trained on COCO2017 train dataset, tested on COCO2017 val dataset.

mAP is IoU=0.5:0.95,area=all,maxDets=100,mAP(COCOeval,stats[0]).

Model	resize-style	input size	batch	epochs	mAP
ResNet50-RetinaNet	YoloStyle-1024	1024x1024	32	13	36.893
ResNet50-FCOS	YoloStyle-1024	1024x1024	32	13	40.155
ResNet50-DETR	YoloStyle-1024	1024x1024	64	500	38.735

You can find more model training details in 03.detection_training/coco/.

All detection models finetune from objects365 pretrain weight on COCO2017

Trained on COCO2017 train dataset, tested on COCO2017 val dataset.

mAP is IoU=0.5:0.95,area=all,maxDets=100,mAP(COCOeval,stats[0]).

Model	resize-style	input size	batch	epochs	mAP
ResNet50-RetinaNet	YoloStyle-1024	1024x1024	32	13	41.259
ResNet50-FCOS	YoloStyle-1024	1024x1024	32	13	45.249

You can find more model training details in 03.detection_training/coco/.

All detection models training from scratch on Objects365(v2,2020)

Trained on objects365(v2,2020) train dataset.

Model	resize-style	input size	batch	epochs	loss
ResNet50-RetinaNet	YoloStyle-1024	1024x1024	128	13	0.3237
ResNet50-FCOS	YoloStyle-1024	1024x1024	128	13	0.9669

You can find more model training details in 03.detection_training/objects365/.

All detection models training from scratch on VOC2007&VOC2012

Trained on VOC2007 trainval dataset + VOC2012 trainval dataset, tested on VOC2007 test dataset.

mAP is IoU=0.50,area=all,maxDets=100,mAP.

Model	resize-style	input size	batch	epochs	mAP
ResNet50-RetinaNet	YoloStyle-640	640x640	32	13	83.460
ResNet50-FCOS	YoloStyle-640	640x640	32	13	83.320

You can find more model training details in 03.detection_training/voc/.

All detection models finetune from objects365 pretrain weight on VOC2007&VOC2012

Trained on VOC2007 trainval dataset + VOC2012 trainval dataset, tested on VOC2007 test dataset.

mAP is IoU=0.50,area=all,maxDets=100,mAP.

Model	resize-style	input size	batch	epochs	mAP
ResNet50-RetinaNet	YoloStyle-640	640x640	32	13	90.034
ResNet50-FCOS	YoloStyle-640	640x640	32	13	89.900

You can find more model training details in 03.detection_training/voc/.

04.semantic_segmentation_training results

pfan_semantic_segmentation

Paper1:https://arxiv.org/abs/1903.00179

Paper2:https://arxiv.org/abs/2210.13452

Paper3:https://arxiv.org/abs/2508.10104

Use ADE20K and COCO2017 dataset to train and test.

Model	dataset	input size	batch	epochs	mean_iou
resnet50_pfan_semantic_segmentation	ADE20K	512x512	32	100	30.326
convformerm36_pfan_semantic_segmentation	ADE20K	512x512	32	100	40.281
dinov3_vit_base_patch16_pfan_semantic_segmentation	ADE20K	512x512	32	100	45.964
resnet50_pfan_semantic_segmentation	COCO2017	512x512	64	100	53.238
convformerm36_pfan_semantic_segmentation	COCO2017	512x512	64	100	61.187
dinov3_vit_base_patch16_pfan_semantic_segmentation	COCO2017	512x512	64	100	64.774

You can find more model training details in 04.semantic_segmentation_training/.

05.instance_segmentation_training results

SOLOv2

Paper:https://arxiv.org/abs/2003.10152

YOLACT

Paper:https://arxiv.org/abs/1904.02689

Trained on COCO2017 train dataset, tested on COCO2017 val dataset.

mAP is IoU=0.5:0.95,area=all,maxDets=100,mAP(COCOeval,stats[0]).

Model	resize-style	input size	batch	epochs	mAP
resnet50_yolact	YoloStyle-1024	1024x1024	64	39	29.211
convformerm36_yolact	YoloStyle-1024	1024x1024	64	39	33.046
dinov3_vit_base_patch16_yolact	YoloStyle-1024	1024x1024	64	39	36.085
resnet50_solov2	YoloStyle-1024	1024x1024	32	39	37.661
convformerm36_solov2	YoloStyle-1024	1024x1024	32	39	40.501
dinov3_vit_base_patch16_solov2	YoloStyle-1024	1024x1024	32	39	43.591

You can find more model training details in 05.instance_segmentation_training/coco/.

06.salient_object_detection_training results

pfan_segmentation

Paper1:https://arxiv.org/abs/1903.00179

Paper2:https://arxiv.org/abs/2210.13452

Paper3:https://arxiv.org/abs/2508.10104

Use combine dataset to train and test.

Model	input size	batch	epochs	iou	precision	recall	f_squared_beta
resnet50_pfan_segmentation	1024x1024	64	100	0.8444	0.8954	0.9335	0.9039
convformerm36_pfan_segmentation	1024x1024	64	100	0.8916	0.9290	0.9549	0.9348
dinov3_vit_base_patch16_pfan_segmentation	1024x1024	64	100	0.9065	0.9439	0.9566	0.9467

You can find more model training details in 06.salient_object_detection_training/.

07.human_matting_training results

pfan_matting

Paper1:https://arxiv.org/abs/1903.00179

Paper2:https://arxiv.org/abs/2104.14222

Paper3:https://arxiv.org/abs/2210.13452

Paper4:https://arxiv.org/abs/2508.10104

Use combine dataset to train and test.

Model	input size	batch	epochs	iou	precision	recall	sad	mae	mse	grad	conn
resnet50_pfan_matting	1024x1024	32	100	0.9823	0.9874	0.9948	6.5496	0.0062	0.0040	10.7192	6.5801
convformerm36_pfan_matting	1024x1024	32	100	0.9881	0.9910	0.9970	4.4842	0.0042	0.0022	8.0214	4.4843
dinov3_vit_base_patch16_pfan_matting	1024x1024	32	100	0.9871	0.9914	0.9955	5.0023	0.0047	0.0026	8.7974	5.0621

You can find more model training details in 07.human_matting_training/.

08.ocr_text_detection_training results

DBNet

Paper:https://arxiv.org/abs/1911.08947

Use combine dataset to train and test.

Model	input size	batch	epochs	precision	recall	f1
resnet50_dbnet	1024x1024	64	100	92.3463	87.1304	89.6626
convformerm36_dbnet	1024x1024	64	100	93.1819	89.5183	91.3134

You can find more model training details in 08.ocr_text_detection_training/.

09.ocr_text_recognition_training results

CTC_Model

Paper:https://arxiv.org/abs/1507.05717

Use combine dataset to train and test.

Model	input size	batch	epochs	lcs_precision	lcs_recall
resnet50_ctc_model	32x512	1024	50	99.1379	98.8073
convformerm36_ctc_model	32x512	1024	50	99.4651	99.2434

You can find more model training details in 09.ocr_text_recognition_training/.

10.face_detection_training results

RetinaFace

Paper:https://arxiv.org/abs/1905.00641

Use combine dataset to train and test.

Model	input size	batch	epochs	Easy AP	Medium AP	Hard AP
resnet50_retinaface	1024x1024	16	100	0.9375	0.9148	0.7804

You can find more model training details in 10.face_detection_training/.

11.face_parsing_training results

pfan_face_parsing

Paper1:https://arxiv.org/abs/1903.00179

Paper2:https://arxiv.org/abs/2210.13452

Paper3:https://arxiv.org/abs/2508.10104

Use CelebAMask-HQ and FaceSynthetics dataset to train and test.

Model	dataset	input size	batch	epochs	precision	recall	iou	dice
resnet50_pfan_face_parsing	CelebAMask-HQ	512x512	192	100	81.4427	77.5129	68.9136	79.2088
convformerm36_pfan_face_parsing	CelebAMask-HQ	512x512	192	100	84.2701	81.5477	72.9179	82.7265
dinov3_vit_base_patch16_pfan_face_parsing	CelebAMask-HQ	512x512	192	100	86.1822	83.7555	75.3506	84.8245
resnet50_pfan_face_parsing	FaceSynthetics	512x512	192	100	95.3781	95.1519	91.2068	95.2643
convformerm36_pfan_face_parsing	FaceSynthetics	512x512	192	100	96.2706	96.1944	92.9115	96.2323
dinov3_vit_base_patch16_pfan_face_parsing	FaceSynthetics	512x512	192	100	95.9629	95.7920	92.2981	95.8769

You can find more model training details in 11.face_parsing_training/.

12.human_parsing_training results

pfan_human_parsing

Paper1:https://arxiv.org/abs/1903.00179

Paper2:https://arxiv.org/abs/2210.13452

Paper3:https://arxiv.org/abs/2508.10104

Use CIHP and LIP dataset to train and test.

Model	dataset	input size	batch	epochs	precision	recall	iou	dice
resnet50_pfan_human_parsing	CIHP	512x512	192	100	62.2381	56.0526	45.2076	58.4858
convformerm36_pfan_human_parsing	CIHP	512x512	192	100	67.9648	63.0336	51.5180	65.0746
dinov3_vit_base_patch16_pfan_human_parsing	CIHP	512x512	192	100	73.1447	70.3957	58.2466	71.6496
resnet50_pfan_human_parsing	LIP	512x512	192	100	56.1325	50.6626	38.8464	52.7264
convformerm36_pfan_human_parsing	LIP	512x512	192	100	61.5202	57.3418	44.6563	59.0827
dinov3_vit_base_patch16_pfan_human_parsing	LIP	512x512	192	100	64.7237	62.3281	48.8433	63.3994

You can find more model training details in 12.human_parsing_training/.

13.interactive_segmentation_training results

SAM

Paper1:https://arxiv.org/abs/2304.02643

Paper2:https://arxiv.org/abs/2508.10104

Use combine dataset to train and test.

You can find all jupyter example in 13.interactive_segmentation_training/sam_predict_example/.

Model	input size	batch	epochs	loss
sam_h_encoder_distill_dinov3_vit_base_patch16_encoder	1024x1024	128	5	0.0013
sam_b	1024x1024	160	2	0.0954
sam_b_multilevel	1024x1024	160	2	0.1413

You can find more model training details in 13.interactive_segmentation_training/.

14.video_interactive_segmentation_training results

SAM2

Paper1:https://arxiv.org/abs/2408.00714

Paper2:https://arxiv.org/abs/2508.10104

Use combine dataset to train and test.

You can find all jupyter example in 14.video_interactive_segmentation_training/sam2_predict_example/.

Model	input size	batch	frame_num	epochs	loss
hiera_l_encoder_distill_dinov3_vit_base_patch16_encoder	1024x1024	24	8	20	0.0438
hiera_b_plus_sam2video_stage1	1024x1024	160	1	2	0.1315
hiera_b_plus_sam2video_stage2	1024x1024	16	8	40	0.4212
hiera_b_plus_sam2video_stage3	1024x1024	16	16	20	0.9382
hiera_b_plus_sam2video_multilevel_stage1	1024x1024	160	1	2	0.1839
hiera_b_plus_sam2video_multilevel_stage2	1024x1024	16	8	40	0.5131
hiera_b_plus_sam2video_multilevel_stage3	1024x1024	16	16	20	0.9516

You can find more model training details in 14.video_interactive_segmentation_training/.

16.universal_segmentation_training

universal_segmentation

Paper:https://arxiv.org/abs/2503.19108

universal_segmentation semantic_segmentation_training results

Model	dataset	input size	batch	epochs	mean_iou
dinov3_vit_large_patch16_universal_segmentation	ADE20K	512x512	128	100	47.8155
dinov3_vit_large_patch16_universal_segmentation	COCO2017	512x512	256	100	64.7959

You can find more model training details in 16.universal_segmentation_training/16.0.semantic_segmentation_training/.

universal_segmentation instance_segmentation_training results

Model	dataset	resize-style	input size	batch	epochs	mAP
dinov3_vit_large_patch16_universal_segmentation	COCO2017	YoloStyle-1024	1024x1024	64	50	45.3113

You can find more model training details in 16.universal_segmentation_training/16.1.instance_segmentation_training/.

universal_segmentation salient_object_detection_training results

Model	input size	batch	epochs	iou	precision	recall	f_squared_beta
dinov3_vit_large_patch16_universal_segmentation	1024x1024	64	50	0.9079	0.9369	0.9651	0.9432

You can find more model training details in 16.universal_segmentation_training/16.2.salient_object_detection_training/.

universal_matting human_matting_training results

Model	input size	batch	epochs	iou	precision	recall	sad	mae	mse	grad	conn
dinov3_vit_large_patch16_universal_matting	1024x1024	32	50	0.9886	0.9913	0.9973	4.1426	0.0039	0.0018	7.7149	4.1218

You can find more model training details in 16.universal_segmentation_training/16.3.human_matting_training/.

universal_matting human_instance_matting_training results

Model	input size	batch	epochs	loss
dinov3_vit_large_patch16_universal_matting	1024x1024	32	50	0.0746

You can find more model training details in 16.universal_segmentation_training/16.4.human_instance_matting_training/.

universal_segmentation face_parsing_training results

Model	dataset	input size	batch	epochs	precision	recall	iou	dice
dinov3_vit_large_patch16_universal_segmentation	CelebAMask-HQ	512x512	256	100	86.6002	84.5362	76.0747	85.5090
dinov3_vit_large_patch16_universal_segmentation	FaceSynthetics	512x512	256	100	97.3316	97.2978	94.8875	97.3139

You can find more model training details in 16.universal_segmentation_training/16.5.face_parsing_training/.

universal_segmentation human_parsing_training results

Model	dataset	input size	batch	epochs	precision	recall	iou	dice
dinov3_vit_large_patch16_universal_segmentation	CIHP	512x512	256	100	80.6561	77.1104	66.0162	78.7259
dinov3_vit_large_patch16_universal_segmentation	LIP	512x512	256	100	67.2514	64.3616	50.9822	65.6268

You can find more model training details in 16.universal_segmentation_training/16.6.human_parsing_training/.