TRAM-V3

Reference checkpoint for TRAM-V3 from the paper TRAM: Transformer-Based Mask R-CNN Framework for Underwater Object Detection in Side-Scan Sonar Data (Makam, Sundaram, & Sundaram).

Architecture: MST (Swin-Base) + FPN + CBAM + CLAHE + Mask R-CNN
Input: 224×224 RGB
Training data: SSS_OD-5 (SeabedObjects-KLSG-derived, plane + ship)
Best epoch: 40 / 60
Random seed: 42 (cuDNN deterministic)

Validation metrics (this run)

Metric	This run	Paper
Det mAP@0.5	0.8767	0.8896
Det mAP@0.5:0.95	0.5666	0.5547
Seg mAP@0.5	0.8461	0.7626
Seg mAP@0.5:0.95	0.4219	0.4234

Usage

git clone -b final-tram-v123 https://github.com/CrypticCortex/iisc-sss-codes.git
cd iisc-sss-codes
pip install -r final/requirements.txt

python -m final.tram_v3.inference \
    --weights /path/to/tram_v3_best.pth \
    --data-root /path/to/SSS_OD-5/valid \
    --output-dir runs/tram-v3/inference \
    --evaluate-map

Files

tram_v3_best.pth — best checkpoint (highest val bbox mAP@0.5:0.95)
training.log — per-epoch training/validation log

Extra checkpoint

tram_v3_best_unseeded.pth is an earlier non-seeded run (best epoch 45 of 65). Reported for comparison; the canonical V3 weights are tram_v3_best.pth from the seeded run.

Source

Code: https://github.com/CrypticCortex/iisc-sss-codes/tree/final-tram-v123/final
Other variants: tram-v1 · tram-v2 · tram-v3

Downloads last month: -; Downloads are not tracked for this model. How to track