4️⃣ Use a Stronger Pretrained Backbone

[Microcosmos22]

Your MobileNet-based custom backbone has no ImageNet weights right now:

torchvision.models.mobilenet_v3_small(pretrained=False)

Switch to:

torchvision.models.mobilenet_v3_small(weights="DEFAULT")

or even better:
💪 resnet50_fpn backbone (default for Mask R-CNN)
→ larger, pretrained feature extractor → huge boost.