rts-ieper/efficientdet-lite0-coco

EfficientDet-Lite0 (COCO)

Detects the 80 COCO object classes (people, vehicles, animals and everyday objects) in the camera image. The model is EfficientDet-Lite0, quantised to int8 with a uint8 320x320 input and the detection post-processing built in. It is more accurate than SSD MobileNet v2 and somewhat slower.

Source

• Model: EfficientDet-Lite0 from google/automl (https://github.com/google/automl/tree/master/efficientdet), trained on COCO 2017.
• ARTPEC-9 and CPU: the Coral test_data file efficientdet_lite0_320_ptq.tflite (https://github.com/google-coral/test_data), used unchanged. It is quantised per channel.
• ARTPEC-8: the official automl efficientdet-lite0 COCO checkpoint, exported by RTS with the same post-processing (25 detections, IoU 0.5) and quantised per tensor, because ARTPEC-8 runs per-tensor models best. The conversion is reproducible (scripts/models/convert/efficientdet-lite0/build.sh) and calibrated on 200 COCO val2017 images.
• Labels: the Coral coco_labels.txt (the same labels as the built-in SSD MobileNet v2 model).

Licence and data

• google/automl is licensed under Apache-2.0. The checkpoint is published by the automl repository and is taken to be under the same licence.
• The model was trained on COCO. The COCO annotations are licensed under CC BY 4.0. COCO images were used for calibration only and are not part of the package.

Accuracy

• Upstream: about 30.4 mAP on COCO for the Coral int8 model.
• Measured by RTS on 50 COCO val2017 images (mAP@0.5:0.95): Coral per-channel model 0.370, our per-tensor ARTPEC-8 model 0.326.

COCO images are not security camera images, so results on a camera will differ.

On an ARTPEC-7 camera

Measured by RTS on an AXIS Q1798-LE (ARTPEC-7 GPU) looking down a street:

• About 824 ms per frame (about 1.2 frames per second), against about 300 ms for the built-in SSD MobileNet v2.
• Scores are lower than with the built-in SSD: at the same thresholds (0.5 to 0.6) it reported far fewer objects. Lower the class thresholds to around 0.4 when you use it.
• SUVs and vans are often classed as "truck". Rules and counters for vehicles should include both "car" and "truck".

On ARTPEC-7 cameras the built-in SSD MobileNet v2 is the better choice. This model is aimed at ARTPEC-8 and ARTPEC-9 cameras, where it should run in tens of milliseconds; it has not been measured there yet.

Limitations

• ARTPEC-8 variant: some post-processing likely runs on the CPU. It has not yet been measured on an ARTPEC-8 camera.
• Small or distant objects are missed more often than large ones.

Variants and speed

Variant	Inference time
artpec9	not measured
artpec8	not measured
cpu	not measured
artpec7 (GPU)	824 ms

Install

Needs EdgeRules 0.5.0 or newer. In EdgeRules open Detection, then Add a model, then RTS catalog, and install EfficientDet-Lite0 (COCO).

Package

• File: efficientdet-lite0-coco-v1.ermodel (8417280 bytes)
• SHA-256: 8b4b7a5c6153fd684e92236bd34b11b717b146bcf9db1eb882894e8dd5ee33c8