PrototypeCommon
ChandlerIdeaCreator/AIC-Webly-supervised-Fine-Grained-Image-Recognition-National-Second-Prize
框架核心是两阶段“粗筛-精滤”数据清洗流程。首先,利用CLIP的多门控决策逻辑进行宏观粗筛,精准剔除插画、图表等非摄影类噪声。随后,利用DINOv2的细粒度特征,创新采用“相对边际分数”识别处于类别边界的混淆样本,并结合GMM模型为各类别动态确定清洗标准。整个流程内置最小样本保留机制,以保护长尾分布下的稀有类别。 在清洗后的高纯度数据上,我们采用非对称协同训练策略。通过Swin Transformer与ConvNeXt双网络,利用其结构差异形成互补的错误识别模式,高效实现相互监督与纠错。同时,我们创新地对模型达成共识的样本目标化应用Balanced Softmax Loss,在精准应对长尾分…
- Owner
- ChandlerIdeaCreator (personal account)
- Stars
- Stars: 210
- Forks
- Forks: 34
- Language
- Python
- License
- None
- Created
- 2025
- Tier
- Prototype (100 to 249 ★) What this tier means
Nearby in Python
All Python repos- agentii-ai/agentii-investment-intelligenceStars: 210PythonPrototype
- evo-design/proto-languageStars: 210PythonPrototype
- DAGroup-PKU/ResearchBoost-AgentStars: 210PythonPrototype
- rmcauley/rainwaveStars: 210PythonPrototype
- Vim-cn/elimageStars: 210PythonPrototype
- openstack/pbrStars: 210PythonPrototype