Clash of Backbones

Which Pytorch Backbone to Use for Low Data Fine-tuning?

Resource-efficient Image Classification

Criteria for choosing backbones for our experiments?

• ImageNet-1k pre-trained weights available in torchvision or github (WaveMix)
• Number of parameters less than 30 M
• Only model with highest ImageNet performance from one architecture family

Backbones Compared

Architecture	# Params (M)	ImageNet-1k Top-1 Accuracy (%)
ResNet-50	25.6	76.13
WaveMix	27.9	75.32
ConvNeXt-Tiny	28.6	82.52
Swin-Tiny	28.3	81.47
SwinV2-Tiny	28.4	82.07
EfficientNetV2-S	21.5	84.23
DenseNet-161	28.7	77.14
MobileNetV3-Large	5.5	75.27
RegNetY-3.2GF	19.4	81.98
ResNeXt-50 32×4d	25.0	81.20
ShuffleNetV2 2.0×	7.4	76.23

Datasets used for benchmarking

Dataset	Domain	# Training Images	# Testing Images	# Classes
CIFAR-10	🖼️ Natural Images	50,000	10,000	10
CIFAR-100	🖼️ Natural Images	50,000	10,000	100
TinyImageNet	🖼️ Natural Images (ImageNet subset)	100,000	10,000	200
Stanford Dogs	🖼️ Natural Images (Dog breeds)	12,000	8,580	120
Flowers-102	🖼️ Natural Images (Flower species)	2,040	6,149	102
CUB-200-2011	🖼️ Natural Images (Bird species)	5,994	5,794	200
Stanford Cars	🖼️ Natural Images (Car models)	8,144	8,041	196
Food-101	🖼️ Natural Images (Food categories)	75,750	25,250	101
DTD	🎨 Texture Images	1,880	1,880	47
UCMerced Land Use	🛰️ Remote Sensing Images	1,680	420	21
EuroSAT	🛰️ Remote Sensing Images	18,900	8,100	10
PlantVillage	🌿 Plant Images	44,343	11,105	39
PlantCLEF	🌿 Plant Images	10,455	1,135	20
Galaxy10 DECals	🌌 Astronomy Images (Galaxy Morphology)	15,962	1,774	10
BreakHis 40×	🏥 Medical Images (Histopathology)	1,398	606	2
BreakHis 100×	🏥 Medical Images (Histopathology)	1,458	632	2
BreakHis 200×	🏥 Medical Images (Histopathology)	1,411	611	2
BreakHis 400×	🏥 Medical Images (Histopathology)	1,276	553	2
RSNA Pneumonia Detection	🏥 Medical Images (Radiology)	24,181	6046	2

Code to run benchmarking for each dataset

python <dataset.py> -model <backbone> -bs <batch-size> 

If you want to create a train test split

python split.py --input_dir <input folder path> --output_dir <output folder path> --test_size <fraction to be split>  

Cite this paper

@misc{jeevan2024backbone,
      title={Which Backbone to Use: A Resource-efficient Domain Specific Comparison for Computer Vision}, 
      author={Pranav Jeevan and Amit Sethi},
      year={2024},
      eprint={2406.05612},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}