Figure 1: Performance of Tailwind-B0 to Tailwind-B2 On MPS Dataset.
Tailwind：用于加拿大一枝黄花入侵检测的双分支门控跨模态转置注意力骨干网络.
This repository is trimmed to the paper version of Tailwind: the TailwindV6 backbone and the configs used in the paper.
We use MMSegmentation v1.1.2 as the codebase.
For install and data preparation, please refer to the guidelines in MMSegmentation v1.1.2.
An example (works for me): CUDA 12.8 and torch 2.8.0+cu1281
pip install torchvision==0.23.0+cu128
pip install mmcv==2.2.0
pip install -r requirements.txt
pip install -e .
Before training or evaluation, update data_root in configs/_base_/datasets/uavm.py or configs/_base_/datasets/mfnet.py to your local dataset path.
Download trained weights and MPS dataset.
(
Google Drive[coming soon] |
Quark | code：4ifc
)
Example: evaluate Tailwind-B0 on MPS:
python tools/test.py configs\taillwindNet\tailwindv6_b0_1xb8-40k_uavm-512x512.py /path/to/checkpoint_file
Example: train Tailwind-B0 on MPS:
python tools/train.py configs\taillwindNet\tailwindv6_b0_1xb8-40k_uavm-512x512.py
Here is a demo script to test a single image. More details refer to MMSegmentation's Doc.
python demo/image_demo.py ${IMAGE_FILE} ${CONFIG_FILE} ${CHECKPOINT_FILE} [--device ${DEVICE_NAME}] [--out-file ${OUT_FILE}]Example: visualize Tailwind-B0 on MPS:
python demo/image_demo.py demo/demo.png configs\taillwindNet\tailwindv6_b0_1xb8-40k_uavm-512x512.py \
/path/to/checkpoint_file --device cuda:0 --out-file result.png