NSDL detector, LOCUS (34 classes): anonymized detector for double-anonymous review
The detect tool detector behind "every LOCUS tool-taking row (Tables 2 and 3, Fig. 3)". YOLO26-L (Ultralytics 8.4.33,
AGPL-3.0), initialised from COCO weights and fine-tuned on the VIKI-L2 train split (900 images) and the LOCUS train split (6300 renders) only; no evaluation image.
34 classes: fetch, stompy, unitree_h1, panda, unitree_go2, anymal_c, apple, banana, bowl, bread, cabinet, knife, peach, pear, plate, sink, table, toaster, tomato, wooden_cutting_board, bottle, cardboardbox, cup, drawer, fork, fridge, kettle, meat, oven, pizza, pumpkin, scissors, spoon, wine.
Inference settings used in the paper: imgsz 960, conf 0.25, IoU 0.70, max_det 64.
Place the file at checkpoints/yolo26l_locus_full34_v1.pt in the anonymized code repository.
Optimizer state and training paths have been removed; the network is the EMA
used for evaluation. Derived from Ultralytics YOLO26 and therefore released
under AGPL-3.0, for research use during review.
| File | Bytes | sha256 |
|---|---|---|
yolo26l_locus_full34_v1.pt |
53163010 | 8f36cd0709e2d456af338012fa6e12eb92529888e00355fefe6e960d8c3538b2 |
- Downloads last month
- 23