| --- |
| license: mit |
| library_name: pytorch |
| pipeline_tag: object-detection |
| tags: |
| - sar |
| - sentinel-1 |
| - xview3 |
| - maritime |
| - vessel-detection |
| - torchscript |
| --- |
| |
| # VARUNA Detector — Ensemble Juara-1 xView3 (TorchScript) |
|
|
| Bobot deteksi kapal SAR yang dipakai lapisan T1 VARUNA. Berkas `traced_ensemble.jit` |
| adalah hasil trace TorchScript dari ensemble solusi juara-1 kompetisi |
| [xView3](https://iuu.xview.us/). |
|
|
| ## Asal |
|
|
| - Solusi juara-1 xView3 oleh Eugene Khvedchenya (BloodAxe). |
| - Repo asli: <https://github.com/DIUx-xView/xView3_first_place> (lisensi MIT). |
| - Skor: peringkat 1 public LB (aggregate 0.603) dan holdout LB (0.617). |
|
|
| ## Arsitektur |
|
|
| - Ensemble 12 model segmentasi multi-head (objectness, vessel, fishing, length). |
| - Total 320.6M parameter, dibekukan sebagai satu graph TorchScript (`traced_ensemble.jit`). |
| - Input: 2 kanal SAR (`vh`, `vv`), keluaran peta skor + regresi panjang. |
|
|
| ## Cara pakai di pipeline VARUNA |
|
|
| - Tiling scene GRD: ukuran tile 2048 px, stride 1536 px (overlap 512 px). |
| - Presisi inferensi: fp16. |
| - Normalisasi input: SigmoidNormalization dengan midpoint -20 dB, temperature 0.18, |
| diterapkan per kanal `vh`/`vv`. |
| - Ambang keputusan: objectness 0.300, vessel 0.338, fishing 0.350 |
| (konfigurasi `1128_b5_b4_vs2_fliplr_test`). |
|
|
| ```python |
| import torch |
| |
| model = torch.jit.load("traced_ensemble.jit").eval().half().cuda() |
| with torch.inference_mode(): |
| out = model(tile_fp16) # tile_fp16: [B, 2, 2048, 2048] kanal (vh, vv) ternormalisasi |
| ``` |
|
|
| ## Integritas |
|
|
| SHA-256 `traced_ensemble.jit`: |
| `c21d6b4204e2803b56014c9780d8d62190ed81d9c2ec56d0cbf437b439307d38` |
|
|
| ## Lisensi dan atribusi |
|
|
| MIT, mengikuti repo asli. Atribusi: Eugene Khvedchenya (BloodAxe), |
| xView3 first place solution, DIUx-xView. Bobot ini didistribusikan ulang |
| tanpa modifikasi arsitektur; hanya dikemas ulang sebagai TorchScript untuk |
| inferensi produksi VARUNA (Finerium, Datathon 2026). |
|
|