Submission: CNOS.v2/TUD-L/cnosv2-tudl

Download submission
Submission name cnosv2-tudl
Submission time (UTC) Nov. 10, 2026, 3:15 a.m.
User jzhao12
Task Model-based 2D segmentation of unseen objects
Dataset TUD-L
Description
Evaluation scores
AP:0.466
AP50:0.759
AP75:0.526
AP_large:0.638
AP_medium:0.460
AP_small:0.000
AR1:0.506
AR10:0.524
AR100:0.524
AR_large:0.680
AR_medium:0.507
AR_small:0.000
average_time_per_image:1.165

Method: CNOS.v2

User jzhao12
Publication
Implementation https://github.com/JZhao12/cnosv2
Training image modalities None
Test image modalities RGB
Description

CNOSv2: CAD-based Novel Object Segmentation with SAM2 + DINOv3.

Submitted to: BOP Challenge 2025, model-based 2D detection and 2D segmentation of unseen objects, BOP-Classic-Core (LM-O, T-LESS, TUD-L, IC-BIN, ITODD, HB, YCB-V), BOP'19 test targets.

Training data: none. CNOSv2 is zero-shot: SAM 2.1 Hiera-L and DINOv3 ViT-L/16 are used with their public pretrained weights (sam2.1_hiera_large.pt, dinov3_vitl16_pretrain_lvd1689m), no fine-tuning on BOP data.

Onboarding data: 42 templates per object rendered with pyrender from the CAD model (viewpoints of the level-0 icosphere, camera distance 0.4 of the predefined poses, black background, uniform grey material for T-LESS). No real onboarding images.

Used 3D models: the default BOP CAD models (models/, models_cad/ for T-LESS).

Notes: the three-stage CNOS pipeline (Nguyen et al., ICCV-W 2023) with newer backbones. (1) Proposals: SAM 2.1 automatic mask generator on the RGB image resized to 640 px width, points_per_side 32, pred_iou_thresh 0.8, stability_score_thresh 0.85 (the SAM v1 value 0.97 is too strict for SAM2 and discards most proposals), box_nms_thresh 0.7; proposals smaller than 5% of the image side or 0.03% of the image area are dropped. (2) Descriptors: each proposal is masked, cropped, padded to a square and resized to 224x224; the DINOv3 CLS token (1024-d) is the descriptor, the same for the templates. (3) Matching: cosine similarity between a proposal and the 42 templates of every object, aggregated as the mean of the top 5; the object with the highest score is assigned, proposals below 0.15 are rejected, per-object NMS at IoU 0.25, at most 100 instances per image. The same prediction file serves both tasks (modal mask as RLE for segmentation, amodal box for detection); time is the wall time of stages 1 to 3 per image. Inference runs in bf16-mixed with pytorch-lightning.

Computer specifications Intel Core Ultra 9 285K, 24 cores; NVIDIA GeForce RTX 5070, 12 GB; Ubuntu 24.04.5 LTS, Linux 7.0.0-42-generic