FGIR-Backbones: backbones × training and evaluation settings on 17 fine-grained datasets

These are the checkpoints behind A Large-Scale Study on the Accuracy vs Cost Trade-offs of Training and Evaluation Settings in Fine-Grained Image Recognition (arXiv:2605.18700), presented at the FGVC13 workshop at CVPR 2026. The study trains 16 ImageNet-pretrained backbones (CNNs and vision transformers) on 17 fine-grained image recognition (FGIR) datasets under four training strategies, plus a 448 px set of runs, and compares accuracy against training and inference cost. Training code: arkel23/FGIR-Backbones. Loader: arkel23/fgir-zoo.

825 checkpoints, one per (dataset, backbone, strategy, image size), each from one training seed. The collection FGIR-Backbones (FGVC13 @ CVPR 2026) groups this repo with the paper.

Strategies

Strategy	Meaning
ft	full fine-tune
fz	frozen backbone, linear head
cal	CAL (counterfactual attention learning)
cal_cm	CALMix (CAL with cross-image discriminative-region mixing)

CAL-NC and CALMix-NC, the paper's no-crop evaluation variants, are not separate files. They evaluate a cal or cal_cm checkpoint without the second forward pass on the attention-guided crop (cal_ap_only=True in the config).

Layout

One folder per strategy and image size, files named {dataset}_{model_name}_{strategy}.pth. manifest.csv lists every file with its dataset, backbone, strategy, image size, training serial and seed, class count, accuracy, SHA-256 and size.

Folder	Strategy	Image size	Serial	Backbones	Datasets	Files
ft_224	ft	224	1, 5	16	17	181
fz_224	fz	224	1, 5	16	17	181
cal_224	cal	224	1, 5	16	17	181
cal_cm_224	cal_cm	224	8	16	4	64
ft_384	ft	384	3, 6	16	4	64
fz_384	fz	384	3, 6	16	4	64
cal_384	cal	384	3, 6	16	4	64
cal_448	cal	448	15	5	3	13
cal_cm_448	cal_cm	448	11	5	3	13

Serial is the run group in the paper's experiment log. The 9 core backbones (VGG-19, ResNet-101, ResNetV2-101, BiT-M ResNetV2-101x3, ViT-B/16, BEiTv2-B/16, Swin-B, ConvNeXt-B, VAN-B3) are trained on all 17 datasets at 224 px; the other 7 on aircraft, cub, soygene and soylocal. At 384 px the Swin backbones use their 384 px variants (swin_*_window12_384*). The 448 px folders hold ViT-B/16 and four torchvision ResNets (resnet18, tv_resnet34, tv_resnet50, tv_resnet101) on aircraft, cars and cub.

Each file is a torch.save dict with four keys: config (the full training configuration, an argparse.Namespace), model (the state dict), accuracy and epoch. There is no optimizer state. The config drives the rebuild, so a file loads without the training repository.

Load a checkpoint and classify an image

fgir_zoo holds a frozen copy of the model code and pins timm==0.9.12.

import torch
from PIL import Image
from torchvision import transforms
from fgir_zoo import backbones

model = backbones.create_model('cal_224/cub_vit_b16_cal')  # or model_name='vit_b16', dataset='cub', strategy='cal'
cfg = model.config
tf = transforms.Compose([
    transforms.Resize((cfg.test_resize_size, cfg.test_resize_size),
                      interpolation=transforms.InterpolationMode.BICUBIC),
    transforms.CenterCrop(cfg.image_size),
    transforms.ToTensor(),
    transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
# a CUB-200-2011 test image, class index 50 (051.Horned_Grebe)
x = tf(Image.open('Horned_Grebe_0050_34561.jpg').convert('RGB')).unsqueeze(0)
with torch.no_grad():
    logits, _ = model(x)          # CAL returns (logits, attention crops)
    model.model.ap_only = True    # CAL-NC: skip the forward pass on the crop
    logits_nc = model(x)
print(logits.argmax(-1).item(), logits.softmax(-1).max().item())        # 50 0.9896
print(logits_nc.argmax(-1).item(), logits_nc.softmax(-1).max().item())  # 50 0.9902

The raw file is one Hub call away:

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download('ERISLab/FGIR-Backbones', 'cal_224/cub_vit_b16_cal.pth')
ckpt = torch.load(path, map_location='cpu', weights_only=False)
ckpt['config'].model_name, ckpt['accuracy']

Accuracy of the released checkpoints

Top-1 accuracy (%) stored in each file: the run's final accuracy on the dataset's test split, as recorded in the experiment log. Each value is one seed; the paper reports means over two or three seeds, so its tables differ slightly. The 17 files below 10% are all BEiTv2 or soyglobal (1,938 classes) cells that fail to train on every seed; they are the study's results, not damaged files.

ft_224: full fine-tune, 224 px

Backbone	aircraft	cars	cotton	cub	dafb	dogs	flowers	food	inat17	moe	nabirds	pets	soyageing	soygene	soyglobal	soylocal	vegfru
beitv2_base_patch16_224_in22k	85.33	88.84	2.92	86.95	92.56	88.87	98.59	91.49	71.54	94.05	87.90	77.65	43.31	9.42	2.94	14.50	95.95
convnext_base	86.35			81.96										30.32		19.50
convnext_base_in22k	87.49	80.10	37.08	84.52	87.92	86.77	99.46	90.65	67.40	93.24	83.84	92.94	35.86	36.88	2.89	26.17	94.41
convnext_large_in22k	83.74			88.30										33.67		33.83
deit3_base_patch16_224	84.79			81.86										45.42		24.50
deit3_base_patch16_224_in21ft1k	84.82			87.76										22.93		27.50
deit3_large_patch16_224_in21ft1k	85.99			90.30										49.39		20.00
resnet101	81.85	85.97	12.50	76.89	86.52	91.06	90.91	84.49	59.22	89.16	74.86	92.64	53.43	26.64	7.28	16.17	85.07
resnetv2_101	82.60	86.94	16.25	76.91	89.37	90.31	92.36	85.32	59.79	92.80	74.38	92.04	52.93	30.81	8.26	13.67	86.30
resnetv2_101x3_bitm_in21k	86.08	88.57	44.17	88.78	91.56	88.86	99.37	90.61	69.90	95.45	86.20	93.95	68.93	50.79	23.05	38.00	95.69
swin_base_patch4_window7_224	85.42			84.47										25.71		24.67
swin_base_patch4_window7_224_in22k	87.73	90.40	49.17	90.56	92.35	88.22	99.63	92.37	73.54	95.72	88.86	94.74	50.97	31.80	7.41	26.33	96.11
swin_large_patch4_window7_224_in22k	88.45			91.15										36.64		30.33
van_b3	86.47	89.39	41.67	79.32	91.26	95.05	96.05	88.46	64.61	94.67	82.50	95.07	65.54	37.49	8.34	20.83	90.27
vgg19_bn	78.55	86.56	30.42	76.13	85.75	85.58	94.28	81.71	53.77	91.68	72.06	92.29	43.78	53.92	21.26	31.67	82.38
vit_b16	82.12	86.69	50.42	87.83	89.96	90.89	99.32	89.66	66.06	94.90	83.77	93.89	37.25	31.16	18.78	25.83	93.65

fz_224: frozen backbone, linear head, 224 px

Backbone	aircraft	cars	cotton	cub	dafb	dogs	flowers	food	inat17	moe	nabirds	pets	soyageing	soygene	soyglobal	soylocal	vegfru
beitv2_base_patch16_224_in22k	55.36	64.36	29.58	90.49	41.45	89.95	99.48	90.38	66.58	73.80	87.18	93.73	26.16	15.92	6.66	22.50	94.86
convnext_base	53.92			66.36										26.42		17.50
convnext_base_in22k	62.20	67.43	40.00	87.73	55.28	89.07	99.53	90.00	61.27	84.54	81.24	93.89	38.99	32.52	14.93	36.33	95.21
convnext_large_in22k	62.89			89.28										36.77		37.17
deit3_base_patch16_224	61.36			69.81										23.68		21.83
deit3_base_patch16_224_in21ft1k	65.92			80.01										23.82		30.00
deit3_large_patch16_224_in21ft1k	70.87			82.34										26.61		26.17
resnet101	44.07	46.75	20.83	62.96	36.08	87.06	83.74	57.45	28.52	68.84	51.90	90.46	24.12	13.63	6.45	12.83	66.72
resnetv2_101	46.47	47.59	26.25	58.72	32.05	84.59	84.34	58.00	26.34	75.23	46.32	89.29	21.82	16.57	6.91	23.33	64.83
resnetv2_101x3_bitm_in21k	52.15	62.98	37.92	87.26	50.53	89.35	99.27	86.52	56.47	82.40	80.69	92.94	41.90	28.41	16.24	27.50	93.85
swin_base_patch4_window7_224	59.92			76.11										29.97		27.00
swin_base_patch4_window7_224_in22k	68.02	76.20	40.83	91.04	59.38	88.67	99.63	91.11	67.47	86.75	87.48	94.33	38.16	36.60	21.67	35.00	96.08
swin_large_patch4_window7_224_in22k	67.75			90.61										39.50		39.67
van_b3	56.02	61.85	36.25	70.02	41.19	95.34	92.29	70.04	31.61	80.12	58.61	92.70	32.24	21.92	10.47	24.00	76.92
vgg19_bn	47.58	47.28	28.33	62.00	31.83	85.45	86.66	60.66	29.33	76.08	51.45	89.48	32.18	20.68	10.32	24.33	71.06
vit_b16	58.21	62.37	31.67	86.19	65.44	87.75	98.70	82.78	53.24	88.21	79.13	91.47	40.55	17.31	13.52	23.17	90.24

cal_224: CAL (counterfactual attention learning), 224 px

Backbone	aircraft	cars	cotton	cub	dafb	dogs	flowers	food	inat17	moe	nabirds	pets	soyageing	soygene	soyglobal	soylocal	vegfru
beitv2_base_patch16_224_in22k	90.82	92.14	16.67	89.39	94.88	83.29	94.54	92.63	75.43	96.50	89.75	95.04	71.52	69.51	41.38	34.00	93.87
convnext_base	90.46			87.50										64.35		14.67
convnext_base_in22k	93.04	94.44	57.50	91.53	94.67	90.15	99.50	92.76	75.53	96.84	90.59	94.96	82.79	69.12	33.83	36.00	95.90
convnext_large_in22k	92.62			91.47										69.09		37.17
deit3_base_patch16_224	90.40			87.28										73.61		32.50
deit3_base_patch16_224_in21ft1k	91.06			89.18										54.68		34.17
deit3_large_patch16_224_in21ft1k	92.62			91.34										73.77		34.50
resnet101	82.21	89.99	35.42	85.85	89.97	92.17	96.83	87.11	62.52	95.07	84.94	93.92	54.02	31.86	23.89	27.17	91.41
resnetv2_101	81.82	91.74	30.83	85.66	89.41	91.18	96.42	86.85	62.44	95.14	85.25	93.32	48.91	31.23	21.50	25.17	90.47
resnetv2_101x3_bitm_in21k	91.69	94.02	42.92	89.73	94.66	88.68	99.24	90.54	72.25	95.07	88.34	93.89	57.54	72.99	32.97	36.50	94.28
swin_base_patch4_window7_224	91.48			86.40										69.69		20.83
swin_base_patch4_window7_224_in22k	91.96	94.27	42.92	90.82	94.12	88.72	99.63	92.73	76.02	97.04	90.14	95.31	54.67	76.70	50.65	38.00	95.98
swin_large_patch4_window7_224_in22k	92.59			91.51										76.14		39.83
van_b3	92.47	94.49	49.58	88.16	94.78	95.44	97.84	90.32	72.25	96.87	87.61	95.18	69.70	64.96	23.94	27.00	93.26
vgg19_bn	88.75	93.47	42.92	84.31	92.33	86.38	98.08	87.83	62.85	94.90	84.47	92.59	66.65	70.02	20.81	27.17	91.14
vit_b16	85.45	90.55	34.17	88.57	91.82	91.62	99.28	90.65	68.25	96.13	86.31	94.44	36.38	52.10	21.09	25.00	94.19

cal_cm_224: CALMix (CAL with cross-image discriminative-region mixing), 224 px

Backbone	aircraft	cub	soygene	soylocal
beitv2_base_patch16_224_in22k	91.90	91.08	71.72	39.33
convnext_base	91.72	89.26	62.55	41.17
convnext_base_in22k	92.95	92.03	67.65	45.33
convnext_large_in22k	93.40	91.97	70.21	48.67
deit3_base_patch16_224	90.73	87.87	75.62	55.50
deit3_base_patch16_224_in21ft1k	91.15	89.35	66.65	47.17
deit3_large_patch16_224_in21ft1k	93.19	91.42	77.47	52.00
resnet101	86.41	87.54	39.00	24.83
resnetv2_101	85.33	86.57	34.09	26.33
resnetv2_101x3_bitm_in21k	92.02	89.75	70.54	40.83
swin_base_patch4_window7_224	90.91	87.38	67.18	46.33
swin_base_patch4_window7_224_in22k	92.68	91.06	77.65	50.50
swin_large_patch4_window7_224_in22k	92.50	91.44	76.43	52.17
van_b3	92.74	88.94	63.48	37.17
vgg19_bn	92.50	87.25	68.90	45.67
vit_b16	85.87	89.77	58.66	39.67

ft_384: full fine-tune, 384 px

Backbone	aircraft	cub	soygene	soylocal
beitv2_base_patch16_224_in22k	76.66	81.57	7.43	9.33
convnext_base	89.50	83.24	40.16	22.17
convnext_base_in22k	88.57	78.46	40.47	28.67
convnext_large_in22k	84.13	87.66	51.49	32.67
deit3_base_patch16_224	87.04	84.52	62.12	34.50
deit3_base_patch16_224_in21ft1k	87.40	87.69	29.98	31.67
deit3_large_patch16_224_in21ft1k	88.27	91.13	60.82	31.00
resnet101	86.65	80.22	37.72	13.67
resnetv2_101	85.60	80.05	40.22	16.83
resnetv2_101x3_bitm_in21k	89.32	89.77	60.61	42.83
swin_base_patch4_window12_384	88.06	85.35	30.81	25.50
swin_base_patch4_window12_384_in22k	90.31	91.09	44.20	30.83
swin_large_patch4_window12_384_in22k	89.95	92.15	40.74	31.33
van_b3	88.36	75.09	38.92	22.83
vgg19_bn	79.78	75.49	58.06	34.33
vit_b16	85.84	88.99	46.00	25.50

fz_384: frozen backbone, linear head, 384 px

Backbone	aircraft	cub	soygene	soylocal
beitv2_base_patch16_224_in22k	9.03	5.82	1.40	7.67
convnext_base	50.53	57.94	34.19	20.67
convnext_base_in22k	61.63	82.31	39.35	41.33
convnext_large_in22k	61.27	85.83	41.45	40.83
deit3_base_patch16_224	64.00	77.94	36.43	25.50
deit3_base_patch16_224_in21ft1k	67.48	85.57	33.52	29.00
deit3_large_patch16_224_in21ft1k	70.99	87.25	31.80	31.00
resnet101	50.65	66.00	14.43	10.33
resnetv2_101	52.66	60.61	25.80	25.00
resnetv2_101x3_bitm_in21k	57.76	89.25	35.16	30.50
swin_base_patch4_window12_384	62.92	76.96	36.42	30.00
swin_base_patch4_window12_384_in22k	72.10	91.37	41.86	35.50
swin_large_patch4_window12_384_in22k	70.42	90.59	46.40	43.50
van_b3	53.92	58.70	24.25	23.67
vgg19_bn	55.36	62.63	30.77	23.67
vit_b16	49.83	85.52	15.61	22.00

cal_384: CAL (counterfactual attention learning), 384 px

Backbone	aircraft	cub	soygene	soylocal
beitv2_base_patch16_224_in22k	86.77	84.95	75.78	39.50
convnext_base	91.90	90.14	77.50	20.83
convnext_base_in22k	94.87	92.23	80.96	41.33
convnext_large_in22k	93.94	92.46	80.64	42.17
deit3_base_patch16_224	92.65	88.38	80.98	41.00
deit3_base_patch16_224_in21ft1k	93.13	89.56	67.06	39.83
deit3_large_patch16_224_in21ft1k	94.09	91.16	82.95	43.17
resnet101	87.76	89.09	51.21	31.00
resnetv2_101	87.46	88.16	51.13	24.17
resnetv2_101x3_bitm_in21k	92.35	90.87	78.52	35.83
swin_base_patch4_window12_384	93.13	87.61	78.36	19.67
swin_base_patch4_window12_384_in22k	93.49	91.85	83.16	40.00
swin_large_patch4_window12_384_in22k	94.09	91.73	82.78	42.50
van_b3	93.46	89.40	76.44	20.50
vgg19_bn	89.89	86.87	78.43	28.67
vit_b16	89.32	89.80	68.20	14.33

cal_448: CAL (counterfactual attention learning), 448 px

Backbone	aircraft	cars	cub
resnet18	92.41	94.19	87.47
tv_resnet101	94.63	95.01	89.92
tv_resnet34	93.55	94.02	88.44
tv_resnet50	94.54	94.94	89.56
vit_b16			90.71

cal_cm_448: CALMix (CAL with cross-image discriminative-region mixing), 448 px

Backbone	aircraft	cars	cub
resnet18	93.13	93.94	88.18
tv_resnet101	94.45	95.22	90.27
tv_resnet34	93.97	94.79	89.16
tv_resnet50	94.63	94.99	89.63
vit_b16			90.82

Requirements

• fgir-zoo (pip install git+https://github.com/arkel23/fgir-zoo.git), which pins timm==0.9.12
• torch (checked with 2.5.1)

Citation

@inproceedings{rios_large-scale_2026,
  title     = {A Large-Scale Study on the Accuracy vs Cost Trade-offs of Training and Evaluation
               Settings in Fine-Grained Image Recognition},
  author    = {Rios, Edwin Arkel and Surya, Augusto Christian and Gosal, Oswin and Mikael, Fernando and
               Nicole, Mary Madeline and Jang, Kisoon and Lai, Bo-Cheng and Hu, Min-Chun},
  booktitle = {The 13th Workshop on Fine-Grained Visual Categorization (FGVC13) at CVPR 2026},
  note      = {Non-archival extended abstract},
  year      = {2026},
  eprint    = {2605.18700},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi       = {10.48550/arXiv.2605.18700},
  url       = {https://arxiv.org/abs/2605.18700}
}