Hugging Face's logo

ERISLab
/
FGIR-Backbones

Image Classification
PyTorch
timm
fine-grained-image-recognition
counterfactual-attention-learning
Model card Files
xet
 Community
xedwin23x commited on
Commit
f0e5e53
·
verified ·
1 Parent(s): 4dfedda

Add README.md

Browse files
Files changed (1) hide show
  1. README.md +304 -0
README.md ADDED
@@ -0,0 +1,304 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pipeline_tag: image-classification
3
+ library_name: pytorch
4
+ tags:
5
+ - fine-grained-image-recognition
6
+ - image-classification
7
+ - counterfactual-attention-learning
8
+ - timm
9
+ ---
10
+
11
+ # FGIR-Backbones: backbones × training and evaluation settings on 17 fine-grained datasets
12
+
13
+ These are the checkpoints behind *A Large-Scale Study on the Accuracy vs Cost Trade-offs of
14
+ Training and Evaluation Settings in Fine-Grained Image Recognition*
15
+ ([arXiv:2605.18700](https://arxiv.org/abs/2605.18700)), presented at the FGVC13 workshop at CVPR
16
+ 2026. The study trains 16 ImageNet-pretrained backbones (CNNs and vision transformers) on 17
17
+ fine-grained image recognition (FGIR) datasets under four training strategies, plus a 448 px
18
+ set of runs, and compares accuracy against training and inference cost. Training code:
19
+ [arkel23/FGIR-Backbones](https://github.com/arkel23/FGIR-Backbones). Loader:
20
+ [arkel23/fgir-zoo](https://github.com/arkel23/fgir-zoo).
21
+
22
+ 825 checkpoints, one per (dataset, backbone, strategy, image size), each from one training seed.
23
+
24
+ ## Strategies
25
+
26
+ | Strategy | Meaning |
27
+ |---|---|
28
+ | `ft` | full fine-tune |
29
+ | `fz` | frozen backbone, linear head |
30
+ | `cal` | CAL (counterfactual attention learning) |
31
+ | `cal_cm` | CALMix (CAL with cross-image discriminative-region mixing) |
32
+
33
+ CAL-NC and CALMix-NC, the paper's no-crop evaluation variants, are not separate files. They
34
+ evaluate a `cal` or `cal_cm` checkpoint without the second forward pass on the
35
+ attention-guided crop (`cal_ap_only=True` in the config).
36
+
37
+ ## Layout
38
+
39
+ One folder per strategy and image size, files named `{dataset}_{model_name}_{strategy}.pth`.
40
+ `manifest.csv` lists every file with its dataset, backbone, strategy, image size, training serial
41
+ and seed, class count, accuracy, SHA-256 and size.
42
+
43
+ | Folder | Strategy | Image size | Serial | Backbones | Datasets | Files |
44
+ |---|---|---|---|---|---|---|
45
+ | `ft_224` | `ft` | 224 | 1, 5 | 16 | 17 | 181 |
46
+ | `fz_224` | `fz` | 224 | 1, 5 | 16 | 17 | 181 |
47
+ | `cal_224` | `cal` | 224 | 1, 5 | 16 | 17 | 181 |
48
+ | `cal_cm_224` | `cal_cm` | 224 | 8 | 16 | 4 | 64 |
49
+ | `ft_384` | `ft` | 384 | 3, 6 | 16 | 4 | 64 |
50
+ | `fz_384` | `fz` | 384 | 3, 6 | 16 | 4 | 64 |
51
+ | `cal_384` | `cal` | 384 | 3, 6 | 16 | 4 | 64 |
52
+ | `cal_448` | `cal` | 448 | 15 | 5 | 3 | 13 |
53
+ | `cal_cm_448` | `cal_cm` | 448 | 11 | 5 | 3 | 13 |
54
+
55
+ Serial is the run group in the paper's experiment log. The 9 core backbones (VGG-19,
56
+ ResNet-101, ResNetV2-101, BiT-M ResNetV2-101x3, ViT-B/16, BEiTv2-B/16, Swin-B, ConvNeXt-B,
57
+ VAN-B3) are trained on all 17 datasets at 224 px; the other 7 on aircraft, cub, soygene and
58
+ soylocal. At 384 px the Swin backbones use their 384 px variants
59
+ (`swin_*_window12_384*`). The 448 px folders hold ViT-B/16 and four torchvision ResNets
60
+ (`resnet18`, `tv_resnet34`, `tv_resnet50`, `tv_resnet101`) on aircraft, cars and cub.
61
+
62
+ Each file is a `torch.save` dict with four keys: `config` (the full training configuration, an
63
+ `argparse.Namespace`), `model` (the state dict), `accuracy` and `epoch`. There is no optimizer
64
+ state. The config drives the rebuild, so a file loads without the training repository.
65
+
66
+ ## Load a checkpoint and classify an image
67
+
68
+ `fgir_zoo` holds a frozen copy of the model code and pins `timm==0.9.12`.
69
+
70
+ ```python
71
+ import torch
72
+ from PIL import Image
73
+ from torchvision import transforms
74
+ from fgir_zoo import backbones
75
+
76
+ model = backbones.create_backbone('vit_b16', dataset='cub', strategy='cal') # cal_224/cub_vit_b16_cal.pth
77
+ cfg = model.config
78
+ tf = transforms.Compose([
79
+ transforms.Resize((cfg.test_resize_size, cfg.test_resize_size),
80
+ interpolation=transforms.InterpolationMode.BICUBIC),
81
+ transforms.CenterCrop(cfg.image_size),
82
+ transforms.ToTensor(),
83
+ transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
84
+ ])
85
+ # a CUB-200-2011 test image, class index 50 (051.Horned_Grebe)
86
+ x = tf(Image.open('Horned_Grebe_0050_34561.jpg').convert('RGB')).unsqueeze(0)
87
+ with torch.no_grad():
88
+ logits, _ = model(x) # CAL returns (logits, attention crops)
89
+ model.model.ap_only = True # CAL-NC: skip the forward pass on the crop
90
+ logits_nc = model(x)
91
+ print(logits.argmax(-1).item(), logits.softmax(-1).max().item()) # 50 0.9896
92
+ print(logits_nc.argmax(-1).item(), logits_nc.softmax(-1).max().item()) # 50 0.9902
93
+ ```
94
+
95
+ The raw file is one Hub call away:
96
+
97
+ ```python
98
+ import torch
99
+ from huggingface_hub import hf_hub_download
100
+
101
+ path = hf_hub_download('ERISLab/FGIR-Backbones', 'cal_224/cub_vit_b16_cal.pth')
102
+ ckpt = torch.load(path, map_location='cpu', weights_only=False)
103
+ ckpt['config'].model_name, ckpt['accuracy']
104
+ ```
105
+
106
+ ## Accuracy of the released checkpoints
107
+
108
+ Top-1 accuracy (%) stored in each file: the run's final accuracy on the dataset's test split,
109
+ as recorded in the experiment log. Each value is one seed; the paper reports means over two or
110
+ three seeds, so its tables differ slightly. The 17 files below 10% are all BEiTv2 or soyglobal
111
+ (1,938 classes) cells that fail to train on every seed; they are the study's results, not
112
+ damaged files.
113
+
114
+ ### `ft_224`: full fine-tune, 224 px
115
+
116
+ | Backbone | aircraft | cars | cotton | cub | dafb | dogs | flowers | food | inat17 | moe | nabirds | pets | soyageing | soygene | soyglobal | soylocal | vegfru |
117
+ |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
118
+ | `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 |
119
+ | `convnext_base` | 86.35 | | | 81.96 | | | | | | | | | | 30.32 | | 19.50 | |
120
+ | `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 |
121
+ | `convnext_large_in22k` | 83.74 | | | 88.30 | | | | | | | | | | 33.67 | | 33.83 | |
122
+ | `deit3_base_patch16_224` | 84.79 | | | 81.86 | | | | | | | | | | 45.42 | | 24.50 | |
123
+ | `deit3_base_patch16_224_in21ft1k` | 84.82 | | | 87.76 | | | | | | | | | | 22.93 | | 27.50 | |
124
+ | `deit3_large_patch16_224_in21ft1k` | 85.99 | | | 90.30 | | | | | | | | | | 49.39 | | 20.00 | |
125
+ | `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 |
126
+ | `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 |
127
+ | `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 |
128
+ | `swin_base_patch4_window7_224` | 85.42 | | | 84.47 | | | | | | | | | | 25.71 | | 24.67 | |
129
+ | `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 |
130
+ | `swin_large_patch4_window7_224_in22k` | 88.45 | | | 91.15 | | | | | | | | | | 36.64 | | 30.33 | |
131
+ | `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 |
132
+ | `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 |
133
+ | `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 |
134
+
135
+ ### `fz_224`: frozen backbone, linear head, 224 px
136
+
137
+ | Backbone | aircraft | cars | cotton | cub | dafb | dogs | flowers | food | inat17 | moe | nabirds | pets | soyageing | soygene | soyglobal | soylocal | vegfru |
138
+ |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
139
+ | `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 |
140
+ | `convnext_base` | 53.92 | | | 66.36 | | | | | | | | | | 26.42 | | 17.50 | |
141
+ | `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 |
142
+ | `convnext_large_in22k` | 62.89 | | | 89.28 | | | | | | | | | | 36.77 | | 37.17 | |
143
+ | `deit3_base_patch16_224` | 61.36 | | | 69.81 | | | | | | | | | | 23.68 | | 21.83 | |
144
+ | `deit3_base_patch16_224_in21ft1k` | 65.92 | | | 80.01 | | | | | | | | | | 23.82 | | 30.00 | |
145
+ | `deit3_large_patch16_224_in21ft1k` | 70.87 | | | 82.34 | | | | | | | | | | 26.61 | | 26.17 | |
146
+ | `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 |
147
+ | `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 |
148
+ | `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 |
149
+ | `swin_base_patch4_window7_224` | 59.92 | | | 76.11 | | | | | | | | | | 29.97 | | 27.00 | |
150
+ | `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 |
151
+ | `swin_large_patch4_window7_224_in22k` | 67.75 | | | 90.61 | | | | | | | | | | 39.50 | | 39.67 | |
152
+ | `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 |
153
+ | `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 |
154
+ | `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 |
155
+
156
+ ### `cal_224`: CAL (counterfactual attention learning), 224 px
157
+
158
+ | Backbone | aircraft | cars | cotton | cub | dafb | dogs | flowers | food | inat17 | moe | nabirds | pets | soyageing | soygene | soyglobal | soylocal | vegfru |
159
+ |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
160
+ | `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 |
161
+ | `convnext_base` | 90.46 | | | 87.50 | | | | | | | | | | 64.35 | | 14.67 | |
162
+ | `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 |
163
+ | `convnext_large_in22k` | 92.62 | | | 91.47 | | | | | | | | | | 69.09 | | 37.17 | |
164
+ | `deit3_base_patch16_224` | 90.40 | | | 87.28 | | | | | | | | | | 73.61 | | 32.50 | |
165
+ | `deit3_base_patch16_224_in21ft1k` | 91.06 | | | 89.18 | | | | | | | | | | 54.68 | | 34.17 | |
166
+ | `deit3_large_patch16_224_in21ft1k` | 92.62 | | | 91.34 | | | | | | | | | | 73.77 | | 34.50 | |
167
+ | `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 |
168
+ | `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 |
169
+ | `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 |
170
+ | `swin_base_patch4_window7_224` | 91.48 | | | 86.40 | | | | | | | | | | 69.69 | | 20.83 | |
171
+ | `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 |
172
+ | `swin_large_patch4_window7_224_in22k` | 92.59 | | | 91.51 | | | | | | | | | | 76.14 | | 39.83 | |
173
+ | `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 |
174
+ | `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 |
175
+ | `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 |
176
+
177
+ ### `cal_cm_224`: CALMix (CAL with cross-image discriminative-region mixing), 224 px
178
+
179
+ | Backbone | aircraft | cub | soygene | soylocal |
180
+ |---|---|---|---|---|
181
+ | `beitv2_base_patch16_224_in22k` | 91.90 | 91.08 | 71.72 | 39.33 |
182
+ | `convnext_base` | 91.72 | 89.26 | 62.55 | 41.17 |
183
+ | `convnext_base_in22k` | 92.95 | 92.03 | 67.65 | 45.33 |
184
+ | `convnext_large_in22k` | 93.40 | 91.97 | 70.21 | 48.67 |
185
+ | `deit3_base_patch16_224` | 90.73 | 87.87 | 75.62 | 55.50 |
186
+ | `deit3_base_patch16_224_in21ft1k` | 91.15 | 89.35 | 66.65 | 47.17 |
187
+ | `deit3_large_patch16_224_in21ft1k` | 93.19 | 91.42 | 77.47 | 52.00 |
188
+ | `resnet101` | 86.41 | 87.54 | 39.00 | 24.83 |
189
+ | `resnetv2_101` | 85.33 | 86.57 | 34.09 | 26.33 |
190
+ | `resnetv2_101x3_bitm_in21k` | 92.02 | 89.75 | 70.54 | 40.83 |
191
+ | `swin_base_patch4_window7_224` | 90.91 | 87.38 | 67.18 | 46.33 |
192
+ | `swin_base_patch4_window7_224_in22k` | 92.68 | 91.06 | 77.65 | 50.50 |
193
+ | `swin_large_patch4_window7_224_in22k` | 92.50 | 91.44 | 76.43 | 52.17 |
194
+ | `van_b3` | 92.74 | 88.94 | 63.48 | 37.17 |
195
+ | `vgg19_bn` | 92.50 | 87.25 | 68.90 | 45.67 |
196
+ | `vit_b16` | 85.87 | 89.77 | 58.66 | 39.67 |
197
+
198
+ ### `ft_384`: full fine-tune, 384 px
199
+
200
+ | Backbone | aircraft | cub | soygene | soylocal |
201
+ |---|---|---|---|---|
202
+ | `beitv2_base_patch16_224_in22k` | 76.66 | 81.57 | 7.43 | 9.33 |
203
+ | `convnext_base` | 89.50 | 83.24 | 40.16 | 22.17 |
204
+ | `convnext_base_in22k` | 88.57 | 78.46 | 40.47 | 28.67 |
205
+ | `convnext_large_in22k` | 84.13 | 87.66 | 51.49 | 32.67 |
206
+ | `deit3_base_patch16_224` | 87.04 | 84.52 | 62.12 | 34.50 |
207
+ | `deit3_base_patch16_224_in21ft1k` | 87.40 | 87.69 | 29.98 | 31.67 |
208
+ | `deit3_large_patch16_224_in21ft1k` | 88.27 | 91.13 | 60.82 | 31.00 |
209
+ | `resnet101` | 86.65 | 80.22 | 37.72 | 13.67 |
210
+ | `resnetv2_101` | 85.60 | 80.05 | 40.22 | 16.83 |
211
+ | `resnetv2_101x3_bitm_in21k` | 89.32 | 89.77 | 60.61 | 42.83 |
212
+ | `swin_base_patch4_window12_384` | 88.06 | 85.35 | 30.81 | 25.50 |
213
+ | `swin_base_patch4_window12_384_in22k` | 90.31 | 91.09 | 44.20 | 30.83 |
214
+ | `swin_large_patch4_window12_384_in22k` | 89.95 | 92.15 | 40.74 | 31.33 |
215
+ | `van_b3` | 88.36 | 75.09 | 38.92 | 22.83 |
216
+ | `vgg19_bn` | 79.78 | 75.49 | 58.06 | 34.33 |
217
+ | `vit_b16` | 85.84 | 88.99 | 46.00 | 25.50 |
218
+
219
+ ### `fz_384`: frozen backbone, linear head, 384 px
220
+
221
+ | Backbone | aircraft | cub | soygene | soylocal |
222
+ |---|---|---|---|---|
223
+ | `beitv2_base_patch16_224_in22k` | 9.03 | 5.82 | 1.40 | 7.67 |
224
+ | `convnext_base` | 50.53 | 57.94 | 34.19 | 20.67 |
225
+ | `convnext_base_in22k` | 61.63 | 82.31 | 39.35 | 41.33 |
226
+ | `convnext_large_in22k` | 61.27 | 85.83 | 41.45 | 40.83 |
227
+ | `deit3_base_patch16_224` | 64.00 | 77.94 | 36.43 | 25.50 |
228
+ | `deit3_base_patch16_224_in21ft1k` | 67.48 | 85.57 | 33.52 | 29.00 |
229
+ | `deit3_large_patch16_224_in21ft1k` | 70.99 | 87.25 | 31.80 | 31.00 |
230
+ | `resnet101` | 50.65 | 66.00 | 14.43 | 10.33 |
231
+ | `resnetv2_101` | 52.66 | 60.61 | 25.80 | 25.00 |
232
+ | `resnetv2_101x3_bitm_in21k` | 57.76 | 89.25 | 35.16 | 30.50 |
233
+ | `swin_base_patch4_window12_384` | 62.92 | 76.96 | 36.42 | 30.00 |
234
+ | `swin_base_patch4_window12_384_in22k` | 72.10 | 91.37 | 41.86 | 35.50 |
235
+ | `swin_large_patch4_window12_384_in22k` | 70.42 | 90.59 | 46.40 | 43.50 |
236
+ | `van_b3` | 53.92 | 58.70 | 24.25 | 23.67 |
237
+ | `vgg19_bn` | 55.36 | 62.63 | 30.77 | 23.67 |
238
+ | `vit_b16` | 49.83 | 85.52 | 15.61 | 22.00 |
239
+
240
+ ### `cal_384`: CAL (counterfactual attention learning), 384 px
241
+
242
+ | Backbone | aircraft | cub | soygene | soylocal |
243
+ |---|---|---|---|---|
244
+ | `beitv2_base_patch16_224_in22k` | 86.77 | 84.95 | 75.78 | 39.50 |
245
+ | `convnext_base` | 91.90 | 90.14 | 77.50 | 20.83 |
246
+ | `convnext_base_in22k` | 94.87 | 92.23 | 80.96 | 41.33 |
247
+ | `convnext_large_in22k` | 93.94 | 92.46 | 80.64 | 42.17 |
248
+ | `deit3_base_patch16_224` | 92.65 | 88.38 | 80.98 | 41.00 |
249
+ | `deit3_base_patch16_224_in21ft1k` | 93.13 | 89.56 | 67.06 | 39.83 |
250
+ | `deit3_large_patch16_224_in21ft1k` | 94.09 | 91.16 | 82.95 | 43.17 |
251
+ | `resnet101` | 87.76 | 89.09 | 51.21 | 31.00 |
252
+ | `resnetv2_101` | 87.46 | 88.16 | 51.13 | 24.17 |
253
+ | `resnetv2_101x3_bitm_in21k` | 92.35 | 90.87 | 78.52 | 35.83 |
254
+ | `swin_base_patch4_window12_384` | 93.13 | 87.61 | 78.36 | 19.67 |
255
+ | `swin_base_patch4_window12_384_in22k` | 93.49 | 91.85 | 83.16 | 40.00 |
256
+ | `swin_large_patch4_window12_384_in22k` | 94.09 | 91.73 | 82.78 | 42.50 |
257
+ | `van_b3` | 93.46 | 89.40 | 76.44 | 20.50 |
258
+ | `vgg19_bn` | 89.89 | 86.87 | 78.43 | 28.67 |
259
+ | `vit_b16` | 89.32 | 89.80 | 68.20 | 14.33 |
260
+
261
+ ### `cal_448`: CAL (counterfactual attention learning), 448 px
262
+
263
+ | Backbone | aircraft | cars | cub |
264
+ |---|---|---|---|
265
+ | `resnet18` | 92.41 | 94.19 | 87.47 |
266
+ | `tv_resnet101` | 94.63 | 95.01 | 89.92 |
267
+ | `tv_resnet34` | 93.55 | 94.02 | 88.44 |
268
+ | `tv_resnet50` | 94.54 | 94.94 | 89.56 |
269
+ | `vit_b16` | | | 90.71 |
270
+
271
+ ### `cal_cm_448`: CALMix (CAL with cross-image discriminative-region mixing), 448 px
272
+
273
+ | Backbone | aircraft | cars | cub |
274
+ |---|---|---|---|
275
+ | `resnet18` | 93.13 | 93.94 | 88.18 |
276
+ | `tv_resnet101` | 94.45 | 95.22 | 90.27 |
277
+ | `tv_resnet34` | 93.97 | 94.79 | 89.16 |
278
+ | `tv_resnet50` | 94.63 | 94.99 | 89.63 |
279
+ | `vit_b16` | | | 90.82 |
280
+
281
+ ## Requirements
282
+
283
+ - `fgir-zoo` (`pip install git+https://github.com/arkel23/fgir-zoo.git`), which pins
284
+ `timm==0.9.12`
285
+ - `torch` (checked with 2.5.1)
286
+
287
+ ## Citation
288
+
289
+ ```bibtex
290
+ @inproceedings{rios_large-scale_2026,
291
+ title = {A Large-Scale Study on the Accuracy vs Cost Trade-offs of Training and Evaluation
292
+ Settings in Fine-Grained Image Recognition},
293
+ author = {Rios, Edwin Arkel and Surya, Augusto Christian and Gosal, Oswin and Mikael, Fernando and
294
+ Nicole, Mary Madeline and Jang, Kisoon and Lai, Bo-Cheng and Hu, Min-Chun},
295
+ booktitle = {The 13th Workshop on Fine-Grained Visual Categorization (FGVC13) at CVPR 2026},
296
+ note = {Non-archival extended abstract},
297
+ year = {2026},
298
+ eprint = {2605.18700},
299
+ archivePrefix = {arXiv},
300
+ primaryClass = {cs.CV},
301
+ doi = {10.48550/arXiv.2605.18700},
302
+ url = {https://arxiv.org/abs/2605.18700}
303
+ }
304
+ ```