VisionXLab/Point2RBox-v3-jittor

Point2RBox-v3 (Jittor reproduction)

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.

Results (DOTA-v1.0, official test server, mAP50)

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%.

Files

• checkpoints/point2rbox_v3_1x_dota_ckpt_12.pkl — end-to-end model
• checkpoints/rotated_fcos_1x_dota_using_pseudo_ckpt_12.pkl — stage-2 model
• weights/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 logs

Training environment

• Jittor 1.3.8.5, numpy 1.26.4 (pinned — numpy≥2 silently corrupts jt.array inputs), CUDA 11.2 toolchain (g++-10), single A100 80GB, official hyper-parameters verbatim (end-to-end batch=2; stage-2 batch=4; AdamW 5e-5, grad-clip 35, 12 epochs, LinearLR warmup 500 iters + MultiStepLR [8,11]).
• Config parity is locked by a zero-tolerance L0 test suite (17 tests, including the 74-line per-class SAM rule table byte-copied from upstream).

Known differences from the upstream paper run

1. SAM attention bias: the upstream mobile_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.
2. Upstream's copy-paste stage-2 branch is dead code (condition can never be true); reproduced verbatim.
3. num_workers=0 instead of 2 (Jittor multi-process dataloader deadlock); pure infra, no effect on training math.
4. GPU-mode numeric noise vs PyTorch golden is at the cuDNN-TF32 level (~1e-3); CPU-mode strict-fp32 parity is 1e-5 to bit-exact per component.