DanteLiu/DEIMv2-X-pill-needle

DEIMv2-X Pill + Needle Detector

A fine-tuned DEIMv2-X (DINOv3 ViT-S+ backbone) model for counting pills and insulin pen needle caps in images. Designed for iOS deployment.

Model Details

Property	Value
Architecture	DEIMv2-X (50.3M params)
Backbone	DINOv3 ViT-S+ (native, not distilled)
Input size	640 x 640
Classes	pill (0), needle (1)
Framework	DEIMv2

Performance (dataset_406 validation set, 87 images)

Metric	Value
Exact match (count)	100%
Within +/-2	100%
Inference speed	24 ms/img (RTX 5090)
Confidence threshold	0.5

Files

File	Size	Description
best_stg2.pth	781 MB	PyTorch checkpoint (for training / fine-tuning)
deimv2_x_pill.mlpackage/	98 MB	CoreML fp16 (iOS deployment)
deimv2_x_pill_int8.mlpackage/	50 MB	CoreML int8 (iOS lightweight)
metadata.json	-	iOS integration metadata

CoreML Output Format

• labels (1, 300) int32 — class IDs (0=pill, 1=needle)
• boxes (1, 300, 4) float16 — bounding boxes in xyxy pixel coordinates
• scores (1, 300) float16 — confidence scores

iOS (CoreML)

See metadata.json for input/output specifications. Use confidence threshold 0.5 and filter by class ID.

Training

• Pretrained from COCO checkpoint (DEIMv2-X DINOv3)
• Fine-tuned 42 epochs (30 + 12 EMA stage) on 383 training images
• 2-class: pill (images < 343) + needle (images >= 343)
• 50 hard negative background crops, 4x needle oversample
• Batch size 16

License

Apache 2.0