
Exact checkpoint
Selected enum: MaskRCNN_ResNet50_FPN_Weights.COCO_V1 for maskrcnn_resnet50_fpn, not the separate v2 architecture. The original file is maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth.
How it works
A feature pyramid supports objects at different scales. A region-proposal network suggests candidates, and aligned region features feed classification, box refinement and a separate mask branch. Multiple objects of the same category retain separate masks.
Typical input and output
Input is a list of RGB image tensors in the documented range. The model applies its internal resizing; the implementation defaults include an 800-pixel minimum side and 1333-pixel maximum side. Output dictionaries contain boxes, labels, scores and mask probabilities.
Strengths
Box, class and mask outputs are available together. It provides a clear class-aware comparison with the class-agnostic prompted masks of SAM.
GIS and remote-sensing use
After suitable fine-tuning, instance masks can support individual building, crown or asset inventories. Preserve instance IDs during polygonization. Reconcile objects split across overlapping tiles before counting or calculating areas.
Limitations
COCO labels do not include a complete building-footprint or vegetation inventory. Occluded objects and touching instances remain difficult. Thresholding a mask does not turn it into an authoritative geographic boundary.
Training and fine-tuning
Prepare per-object masks, boxes and class labels. Replace both the box predictor and mask predictor for the new class count, including background as expected by the implementation. Check mask-to-box consistency and empty images. Evaluate mask AP, missed objects and duplicate instances on held-out areas.
Framework and hardware
PyTorch and TorchVision. Use the named weights enum and evaluation mode for prediction. A GPU is generally useful for training and larger images; CPU can be used for a limited compatibility trial. Runtime and memory must be measured for the actual image sizes.
Files, license and download
External official download: 178,090,079 bytes exceeds the 100 MiB resource limit. TorchVision also cautions that pretrained weights may carry training-data terms; its code license alone does not resolve all redistribution rights.
Before using a result
- Keep imagery rights, acquisition dates and preprocessing with each experiment.
- Separate locations before tiling to avoid train/test leakage.
- Inspect failures and uncertain outputs alongside the original imagery.
- Preserve coordinate transforms and verify units before exporting a GIS layer.
Validation status: Source provenance and website delivery are checked. GISSchools has not run model inference, training or an accuracy benchmark for this record. Feature images and gallery diagrams are explanatory illustrations, not predictions.
Official references
Examples and images
Downloads and resources
External official download: 178,090,079 bytes exceeds the 100 MiB resource limit. TorchVision also cautions that pretrained weights may carry training-data terms; its code license alone does not resolve all redistribution rights.



Comments
No comments yet. Start the discussion.
Leave a reply
Your email address will not be published. Required fields are marked.