This is a Jittor reproduction of Point2RBox-v3 (point-supervised oriented object detection with SAM-guided pseudo-label refinement, arXiv:2509.26281), built on JDet. Code: https://github.com/mingqian-233/Point2RBox-v3-jittor.
| Model | Paper (PyTorch) | This repo (Jittor) |
|---|---|---|
| Point2RBox-v3 end-to-end (12 ep) | 59.61 | 59.52 |
| Point2RBox-v3 two-stage (rotated-FCOS) | 66.09 | 65.50 |
The end-to-end number is the official DOTA-v1.0 Task1 test-server mAP50
(0.5952330691632053) and passes the project's paper−2.0 acceptance threshold
(57.61). Per-class AP table: TBD.
Local diagnostic only: the norm-eval-correct end-to-end checkpoint scores 66.6056 mAP50 on the unmerged 1024×1024 trainval-patch protocol. The same-machine full upstream PyTorch run scores 65.60, while the published upstream log reports 66.70. This is not the official DOTA test-server protocol; the official server result is reported in the table above.
The completed two-stage checkpoint scores 75.7029 mAP50 on the same local trainval-patch diagnostic and 65.50 mAP50 on the official DOTA-v1.0 Task1 test server. The latter passes the project's paper−2.0 acceptance threshold of 64.09.
The generated stage-2 pseudo boxes were also matched directly against all 245,953 trainval GT boxes. They achieve mean rotated IoU 0.7330, recall@0.5 91.03%, and recall@0.75 56.58%. The corresponding official-code PyTorch v2 baseline is 0.7295 / 91.25% / 55.39%.
checkpoints/point2rbox_v3_1x_dota_ckpt_12.pkl — end-to-end modelcheckpoints/rotated_fcos_1x_dota_using_pseudo_ckpt_12.pkl — stage-2 modelweights/mobile_sam.pkl — MobileSAM converted weights (439 tensors, numpy
pickle; native Jittor port, mask IoU 1.0000 vs PyTorch on CPU fp32)weights/ted.pkl — TED edge detector converted weights (parity 5.7e-6)logs/ — training logsmobile_sam package calls eval()
before load_state_dict, leaving TinyViT's non-persistent attention-bias
cache at its random initialisation — i.e. the upstream training used an
unseeded random attention bias (unreproducible across runs). This repo
loads the weights correctly. Direction of the effect on mAP is unknown
(expected neutral-to-positive). See docs/config_parity.md.num_workers=0 instead of 2 (Jittor multi-process dataloader deadlock);
pure infra, no effect on training math.