v0.5.1

Changes

• New ckpt trained by 3090 (3.4hrs) with Cross-domain Loss: 15.18 -> 6.59!
• Better support of custom datasets.
• Migrate the recommend package management tool from rye to uv.
• Remove the impl of RCCNet.

v0.5.0

Changes

• Introduce circular smooth label from Arbitrary-Oriented Object Detection with Circular Smooth Label (ECCV'20).
• Now you can use the pipe operator | to build dataset.
• Decrease the DEFAULT_CLS_NUM to 128.

v0.4.2

Changes

• fine tuned RotNetR with SGD and OneCycleLR

v0.4.1

New Features

• Now RotNet is identical with d4nst/RotNet and could produce terrific accuracy
• Breakdown data pipeline into small functions

v0.4.0

New Features

• RotNet is aviliable
• Compatible with Google StreetView Dataset
• Introduce tqdm progress bar

v0.3.0

New Features

• Add implement of d4nst/RotNet. But it still needs further tuning.
• Add a server script

v0.2.1

New Features

• It seems to be the first available version.