Instance Segmentation Base Model

• DeepLavV3+

  A robust baseline model for semantic segmentation, DeepLabV3+ with ResNet-50 strikes a good balance between speed and accuracy. Suitable for medium-scale applications where both performance and resource efficiency are important.

• SegFormer (MiT-B0 backbone)

  A transformer-based semantic segmentation model that integrates hierarchical attention mechanisms and lightweight decoders, achieving efficient and accurate segmentation across diverse image resolutions and

  scales

• SegFormer-large (MiT-B4 backbone)

  Offers significantly better accuracy than MiT-B0 by using a deeper backbone. A great choice for scenarios requiring high-quality segmentation results on complex scenes, while still being manageable on standard GPUs (12–16 GB).

• SegFormer-x-large (MiT-B5 backbone)

  The most accurate SegFormer variant, especially effective on a diverse and densely annotated dataset. Best suited for large-scale training or inference tasks where precision is critical. Requires high-end GPUs with larger memory (≥24 GB).

Best Overall Model (Top Performer)

Segformer x large :

• Use when: You have large datasets, and want best accuracy.
• Why: Offers excellent performance across COCO-style benchmarks; optimized for both speed and scale.

Recommendations by Dataset Size

• Large Dataset (≥ 10k images)

  ✅ SeFormer-x large (MiT-B5)

  ✅ SegFormer-large (MiT-B4)

  ✅ DeepLabV3+ (ResNet-101)

• Why: These models scale well with large datasets and offer high accuracy on complex multi-class segmentation like ADE20K.

• Medium Dataset (2k–10k images)

  ✅ SegFormer-large (MiT-B4)

  ✅ DeepLabV3+ (ResNet-50)

• Why: Balanced models offering a good trade-off between accuracy and training time on medium-scale datasets.

• Small Dataset (≤ 2k images)

  ✅ SegFormer (MiT-B0)

  ✅ DeepLabV3+ (ResNet-50)

• Why: These are lightweight and stable models that avoid overfitting and work well with limited data and compute.