BlankSkip: Early-exit Object Detection onboard Nano-drones

Dimensions overall score 7.0

• BlankSkip achieves up to 24% average throughput improvement compared to a static MobileNet-SSD detector on a nano-drone platform.result 95%
  Evidencepartial

  we achieve up to 24% average throughput improvement with a limited 0.015 mean Average Precision (mAP) drop compared to a static MobileNet-SSD detector

  Implicationpartial

  Explicitly stated in the abstract with a clear numeric result.

  Verificationpartial

  partial

• BlankSkip's early-exit mechanism is based on a lightweight binary classifier that identifies frames with no objects of interest.method 90%
  Evidencepartial

  Rather than applying the full detection pipeline to every input, we decompose the problem into two sequential sub-tasks: a lightweight binary classifier, attached at an intermediate layer of the backbone, first determines whether a frame contains any object of interest

  Implicationpartial

  Directly stated in the introduction as the core method.

  Verificationpartial

  partial

• The BlankSkip architecture increases the total parameter count by only 1.1–3.0% over the original 4.67M parameter SSD model.technical 90%
  Evidencepartial

  SSD parameter count by only 1.1–3.0% over the original 4.67M.

  Implicationpartial

  Specific numeric values are provided in the analysis of the architecture.

  Verificationpartial

  partial

• Deploying object detection on nano-drones is complicated by extremely tight constraints of approximately 10 MiB memory and a 1 W power budget.Source prooftechnical 95%
  Evidenceverified

  complicated by the extremely tight constraints of their computational platforms (approximately 10 MiB memory, 1 W power budget).

  Implicationpartial

  Explicitly stated in the abstract as the core challenge.

  Verificationverified

  verified

• The system is trained end-to-end with a composite loss function that includes a weighted binary cross-entropy term for the early-exit branch.method 90%
  Evidencepartial

  The system is trained end-to-end with a composite loss: L_total = L_loc + L_cls + λ · L_EE

  Implicationpartial

  Directly described in the training strategy section with the loss formula provided.

  Verificationpartial

  partial

• Higher early-exit branch weight w_0 penalizes false positives (non-empty images classified as empty), prioritizing detection completeness.technical 85%
  Evidencepartial

  Higher w_0 penalizes false positives (non-empty images incorrectly classified as empty), prioritizing detection completeness

  Implicationpartial

  Explicitly stated in the description of the training loss tuning trade-off.

  Verificationpartial

  partial

• The method was deployed and evaluated on a real-world nano-drone platform, the Bitcraze Crazyflie 2.1 with a GAP8 SoC.result 90%
  Evidencepartial

  With experiments using a real-world nano-drone platform, the Bitcraze Crazyflie 2.1

  Implicationpartial

  Explicitly stated in the abstract and detailed in the platform description.

  Verificationpartial

  partial

• All DNNs were quantized to 8-bit integer format using Quantization-Aware Training (QAT) for deployment on the GAP8 SoC.method 85%
  Evidencepartial

  we quantize all DNNs to 8-bit integer format with Quantization-Aware Training (QAT)

  Implicationpartial

  Directly stated in the deployment methodology section.

  Verificationpartial

  partial

Competitive landscape

Enables real-time object detection on ultra-low-power nano-drones by adaptively skipping computation for frames without objects of interest.