DanteLiu/deimv2-x-pill-android

DEIMv2-X Pill + Needle Detector (Android ONNX, fp16)

A fine-tuned DEIMv2-X object detection model for counting pills and insulin pen needle caps. Exported to ONNX with fp16 weights for efficient Android deployment via ONNX Runtime.

Model Details

Architecture	DEIMv2-X (ViT backbone + Hybrid Encoder + Deformable DETR Decoder)
Parameters	50.3M
Input	[1, 3, 640, 640] float16, NCHW, ImageNet normalized
Classes	pill (0), needle (1)
Framework	ONNX Runtime Android 1.21.0
Precision	fp16 (101 MB)
License	Apache 2.0

Performance

Metric	Value
Exact count match (fp32)	100% (87/87 val images)
Exact count match (fp16)	99% (15/16 test images)
Within ±2	100%
Confidence threshold	0.5
Inference (modern Android)	~50–200ms

Files

File	Size	Description
deimv2_x_pill_fp16.onnx	101 MB	fp16 ONNX model (plaintext)
deimv2_x_pill_fp16.onnx.enc	101 MB	AES-256-CTR encrypted model (for app distribution)
metadata.json	1 KB	Model I/O specification and metadata

Encrypted Model

The .onnx.enc file is encrypted with AES-256-CTR to prevent direct extraction from the APK.

• Format: [16-byte IV][ciphertext]
• Decryption: Handled by ModelDecryptor.kt in the Android app at runtime

Quick Start (Android / Kotlin)

Gradle

dependencies {
    implementation 'com.microsoft.onnxruntime:onnxruntime-android:1.21.0'
}

Preprocessing

1. Resize image to 640 x 640 (stretch, no padding)
2. Scale pixels to [0, 1] (divide by 255)
3. Normalize: (pixel / 255.0 - mean) / std
   - mean = [0.485, 0.456, 0.406]
   - std  = [0.229, 0.224, 0.225]
4. Layout: NCHW, convert to float16

Inference

val ortEnv = OrtEnvironment.getEnvironment()
val session = ortEnv.createSession(modelFilePath, OrtSession.SessionOptions())
val results = session.run(mapOf("image" to inputTensor))

Postprocessing

For each of the 300 detections:
  1. Skip if scores[0][i] <= 0.5
  2. Get label: labels[0][i]  (0 = pill, 1 = needle)
  3. Get box: [x1, y1, x2, y2] = boxes[0][i]  (640x640 space)
  4. Scale to original image:
       x_orig = x / 640.0 * original_width
       y_orig = y / 640.0 * original_height

Model I/O

Input:

Name	Shape	Type	Description
image	[1, 3, 640, 640]	float16	Preprocessed RGB image (NCHW)

Outputs:

Name	Shape	Type	Description
labels	[1, 300]	int32	Class: 0 = pill, 1 = needle
boxes	[1, 300, 4]	float16	Bounding boxes [x1, y1, x2, y2] (0–640)
scores	[1, 300]	float16	Confidence scores (0–1)

Key Notes

• End-to-end: DEIMv2 outputs are final — no NMS or post-processing needed
• fp16 I/O: Use ShortBuffer + OnnxJavaType.FLOAT16 on Android
• 300 queries: Always outputs 300 candidates; most have near-zero scores
• Stream from file: Load model from file path (not readBytes()) to avoid OOM on memory-constrained devices

Known Limitations

1. Blister packs: Reflective foil may cause +1–2 overcounting
2. Very small pills: May be missed if too small relative to image — move camera closer
3. Stacked pills: 3+ layers of stacking may cause occlusion
4. Needle types: Trained on insulin pen needle caps only

Training Data

• 333 pill images + 60 needle images
• Custom dataset, not publicly available

Citation

@misc{reveliorx2026,
  title={RevelioRX: DEIMv2-X Pill and Needle Detector},
  author={Tsung-Han Liu},
  year={2026},
  url={https://huggingface.co/DanteLiu/deimv2-x-pill-android}
}