Ensamble3

Ensamble 3 — SurgVU 2026 Category 1

The three RF-DETR-Medium checkpoints of the ensemble submitted to MICCAI SurgVU 2026, Category 1 (clevis detection and instrument classification in robot-assisted surgical video), where it placed 6th among the competing teams at 0.5385 mAP on the final-phase test set.

Inference code, Dockerfile and full documentation: https://github.com/GiorgioChiesa/VuSurg26_Ensamble3

The checkpoints

All three share one architecture (RF-DETR-Medium, 576x576, 14 classes) and one training schedule. They differ only in the external pre-train of stage 1, which is what makes them useful to ensemble.

Each is 127.7 MB. They are meant to be used together: fused with weighted box fusion at IoU 0.55 plus horizontal-flip test-time augmentation, six forward passes per frame. A single member scores 0.6288 at best; the trio scores 0.6414.

Classes

The index is the class id and the order must not be changed:

0  grasping retractor              7  monopolar curved scissor
1  cadiere forceps                 8  vessel sealer
2  bipolar forceps                 9  tip up fenestrated grasper
3  force bipolar                  10  bipolar dissector
4  clip applier                   11  needle driver
5  stapler                        12  prograsp forceps
6  permanent cautery hook spatula 13  suction irrigator

Measured performance

Seven-fold leave-two-out on the seven released challenge videos, mAP@[.5:.05:.95], class-aware, at the challenge's 640x512 prediction geometry:

Precision is low by design. Average precision integrates the precision-recall curve, so a low-scoring false positive ranked below the true positives costs almost nothing while a discarded true positive costs a whole object. The container therefore emits every surviving detection; a 96-point sweep confirmed that the detector library's default post-processing would cost 0.0396 mAP.

Honest caveat on this number

The 0.6414 figure describes models cross-validated on a 9,814-frame pseudo-label corpus. These weights were trained two days later on 10,486 frames. It is the best available estimate of what is published here, not a measurement of it, and the discrepancy runs in the favourable direction (more data). It is stated because a number carried without its provenance is how a model card becomes wrong.

Training data

Seven annotated challenge videos (11,323 boxes) plus a pseudo-labelled corpus of 10,486 frames / 23,033 boxes grown from unannotated SurgVU footage. The corpus was admitted through two independent gates — a detection-confidence floor of 0.20 and an exact-multiset consistency check against the operating-room record of installed instruments — and 6,569 of its frames were then corrected by hand.

Six of the fourteen categories (stapler, permanent cautery hook spatula, tip-up fenestrated grasper, bipolar dissector, prograsp forceps, suction irrigator) have zero ground-truth boxes in the released annotation. They are scored on the hidden test set, but no locally computable metric says anything about them in either direction. Treat their predictions with corresponding caution.

Usage

The supported path is the container in the GitHub repository, which downloads these files automatically. To fetch them directly:

from huggingface_hub import hf_hub_download

for name in ("clevis_sarrarp.pth", "clevis_mixA.pth", "clevis_mixC.pth"):
    hf_hub_download("SGio16/Ensamble3", name, local_dir="model/ensemble")

Integrity

41d0c9a77a67f960d07515f8a88ba2adcb4a454b1bb10ee02b478808b0b93bdc  clevis_sarrarp.pth
539d70280c4c6462515a0d79e3a0bb2d666f8a83336ae3858dc5ea04bf1db13a  clevis_mixA.pth
33a42a8372fa0ee31e8e8ebc6cc2fd190b3ec9cfb40279b9e78ee1f2e1395637  clevis_mixC.pth

License

Not yet set by the authors. These weights are derived from the SurgVU dataset; any use is subject to the SurgVU data agreement with the challenge organisers.

Citation

• Zia, A., Berniker, M., Nespolo, R., Zhang, X., Perreault, C., Wang, Z., Mueller, B., Schmidt, R., Bhattacharyya, K., Liu, X., & Jarc, A. (2025). Surgical Visual Understanding (SurgVU) Dataset. arXiv:2501.09209
• Robinson, I., Robicheaux, P., Popov, M., Ramanan, D., & Peri, N. (2026). RF-DETR: Neural Architecture Search for Real-Time Detection Transformers. ICLR. arXiv:2511.09554
• Solovyev, R., Wang, W., & Gabruseva, T. (2021). Weighted boxes fusion. Image and Vision Computing, 107, 104117.