About YarnCV
-
acquisition
- camera sampling at fps=30
-
Yarn Dataset
- directory: "data/refined" and "data/refined-2/"; Note: label-3 is sikpped since it depends heavily on real-time;
- num classes=4: 0=normal, 1~4=abnormal.
-
model, ckpt and metrics
- models: modified resnet18, modified efficientnet-b0, yarn_sim, yarn_sim2 (best acc=100)
- ckpt: see "out/ckpt/"
- metrics: accuracy
-
another attempt: visual large model
- challenge: yarn_sim2 has poor acc (~50) on img-test
- prompt engineering:
- Qwen-vl-max: ~75 acc
- future work: supervised fine-tuning
- method: peft + LoRA
- useful platforms: bailian.console.aliyun.com
-
related resources (urls)
- data: see data/.gitkeep
- ckpt: see src/recognition/out/ckpt/.gitkeep