This is the official winning solution of 2025 IEEE Low Power Computer Vision Challenge Track1. (LPCV 2025) LabLVM
LPCV2025 is the 8th workshop on Efficient Deep Learning for Computer Vision at CVPR 2025.
The repository contains code for layer imitation training, evaluation, and submission to AIhub.
- We use MobileCLIP-S1 as the base model for layer imitation.
- Our approach reduces inference time by 13% and the number of parameters by 15%, with only a 1% decrease in accuracy on our custom dataset.
- We do not use any synthetic images; fine-tuning and testing are performed exclusively on the COCO 2014 and COCO 2017 datasets.
Our experiments are conducted using PyTorch 2.3.0 and CUDA 12.1. Install the required packages by running the following command.
pip install -e .
To download the MobileCLIP-S1 weight from ml-mobileclip, run the command below.
wget https://docs-assets.developer.apple.com/ml-research/datasets/mobileclip/mobileclip_s1.pt -P weight
To build text embedding, run the command below.
python extract_text_features.py
Download the COCO datasets and organize them according to the directory structure shown below.
coco
├── train2014
├── val2014
├── test2014
├── train2017
├── val2017
└── test2017
To imitate the original layer, run train.py. The model weight will be saved to output directory.
python train.py
To reproduce the results, run eval.py.
python eval.py
The results are evaluated on a custom dataset composed of COCO 2014 and COCO 2017.
| Parameters (M) | Throughput (ms) | Accuracy (%) | |
|---|---|---|---|
| Baseline | 20.5 | 2.3 | 47.5 |
| Ours | 17.4 | 2.0 | 47.0 |
Run run.py to submit to AIhub.
python run.py