NanoDet: Tiny Object Detection for TFJS and NodeJS

Models included in /model-tfjs-graph-* were converted to TFJS Graph model format from the original repository
Models descriptors and signature have been additionally parsed for readability

Implementation

Actual model parsing implementation in nanodet.js does not follow original Pyhthon implementation and is fully custom and optimized for JavaScript execution

Original model is internally using Int64 values, but TFJS does not support Int64 so there are some overflows due to Int32 casting,
Most commonly around class 62, so that one is excluded from results

Note that NanoDet-G variation is about 4x faster in Browser execution using WebGL backend than NanoDet-M variation

Function processResults() takes output of model.execute and returns array of objects:

• id: internal number of detection box, used only for debugging
• strideSize: internal size of the stride where object was detected, used only for debugging
• score: value 0..1
• class: coco class number
• label: coco label as string
• box: detection box [x1, y1, x2, y2] normalized to input image dimensions
• boxRaw: detection box [x1, y1, x2, y2] normalized to 0..1
• center: center point of detected object [x, y] normalized to input image dimensions
• centerRaw: center point of detected object [x, y] normalized to 0..1

---

Example

---

Conversion Notes

Source: https://github.com/RangiLyu/nanodet

Requirements:

pip install torch onnx onnx-tf onnx-simplifier tensorflowjs

Fixes:

• Error during conversion: pytorch_half_pixel
  Edit export.py to set opset_version=10 which forces onnx export to use older upscale instead of resize op

Conversion:

• From PyTorch to ONNX to TensorFlow Saved model to TensorFlow/JS Graph model

python export.py --cfg_path config/nanodet-m-416.yml --model_path models/nanodet_m_416.pth --out_path models/nanodet_m_416.onnx
python -m onnxsim models/nanodet_m_416.onnx models/nanodet_m_416-simplified.onnx
onnx-tf convert --infile models/nanodet_m_416-simplified.onnx --outdir models/saved-m
tensorflowjs_converter --input_format tf_saved_model --output_format tfjs_graph_model --strip_debug_ops=* --weight_shard_size_bytes 18388608 models/saved-m models/graph-m

---

Tests

node nanodet.js car.jpg

2021-03-16 12:03:39 INFO:  detector version 0.0.1
2021-03-16 12:03:39 INFO:  User: vlado Platform: linux Arch: x64 Node: v15.4.0
2021-03-16 12:03:39 INFO:  Loaded model { modelPath: 'file://models/nanodet/nanodet.json', minScore: 0.15, iouThreshold: 0.1, maxResults: 10, scaleBox: 2.5 } tensors: 524 bytes: 3771112
2021-03-16 12:03:39 INFO:  Loaded image: car.jpg inputShape: [ 2000, 1333, [length]: 2 ] outputShape: [ 1, 3, 416, 416, [length]: 4 ]
2021-03-16 12:03:39 DATA:  Results: [
  {
    score: 0.7859958410263062,
    strideSize: 1,
    class: 3,
    label: 'car',
    center: [ 1000, 1076, [length]: 2 ],
    centerRaw: [ 0.5, 0.8076923076923077, [length]: 2 ],
    box: [ 375, 868, 1625, 1284, [length]: 4 ],
    boxRaw: [ 0.1875, 0.6514423076923077, 0.8125, 0.9639423076923077, [length]: 4 ]
  },
  {
    score: 0.20603930950164795,
    strideSize: 1,
    class: 26,
    label: 'umbrella',
    center: [ 1615, 358, [length]: 2 ],
    centerRaw: [ 0.8076923076923077, 0.2692307692307692, [length]: 2 ],
    box: [ 1302, -57, 1927, 983, [length]: 4 ],
    boxRaw: [ 0.6514423076923077, -0.04326923076923078, 0.9639423076923077, 0.7379807692307692, [length]: 4 ]
  },
  {
    score: 0.16496318578720093,
    strideSize: 4,
    class: 59,
    label: 'potted plant',
    center: [ 865, 858, [length]: 2 ],
    centerRaw: [ 0.4326923076923077, 0.6442307692307693, [length]: 2 ],
    box: [ 748, 754, 943, 910, [length]: 4 ],
    boxRaw: [ 0.3740985576923077, 0.5661057692307693, 0.4717548076923077, 0.6832932692307693, [length]: 4 ]
  },
  {
    score: 0.15522807836532593,
    strideSize: 4,
    class: 14,
    label: 'bench',
    center: [ 1557, 858, [length]: 2 ],
    centerRaw: [ 0.7788461538461539, 0.6442307692307693, [length]: 2 ],
    box: [ 1362, 832, 1753, 884, [length]: 4 ],
    boxRaw: [ 0.6811899038461539, 0.6246995192307693, 0.8765024038461539, 0.6637620192307693, [length]: 4 ]
  },
]
2021-03-16 12:03:39 STATE:  Created output image: car-nanodet.jpg]

---

Notes

• BoxRaw and CenterRaw are normalized to range 0..1
• Box and Center are normalized to input image size in pixels

What's different about this model is that we get 3 different resultsets and need to check each as different strides pick up different sized objects