IEEE Circuits and Systems Magazine - Q3 2023 - 24

The object detection results on Pascal VOC trained
with YOLOv3 [135] are shown in Table 5 including results
for M4 MCU with 256 kB SRAM and H7 MCU with 512 kB
SRAM. On H7 MCU, MCUNet-H7 improves the mAP by
16.7% compared to the state-of-the-art method MCUNet
[8]. It can also scale down to fit a cheaper commodity
Cortex-M4 MCU with only 256 kB SRAM, while still improving
the mAP by 13.2% at 1.9× smaller peak SRAM.
Note that MCUNet-M4 shares a similar computation with
MCUNet (172 M versus 168 M) but a much better mAP.
This is because the expanded search space from patchbased
inference allows us to choose a better configuration
of larger input resolution and smaller models.
Memory-Efficient Face Detection
The performance of MCUNet for memory-efficient face
detection on WIDER FACE [136] dataset are shown in
Table 6. The analytic memory usage of the detector
backbone in fp32 is reported following [105]. The models
are trained with S3FD face detector [140] following
[105] for a fair comparison. The reported mAP is calculated
on samples with ≤3 faces, which is a more realistic
setting for tiny devices. MCUNet outperforms existing
solutions under different scales. MCUNet-L achieves
comparable performance at 3.4× smaller peak memory
compared to RNNPool-Face-C [8] and 9.9× smaller peak
memory compared to LFFD [138]. The computation is
also 1.6× and 8.4× smaller. MCUNet-S consistently outperforms
RNNPool-Face-A [105] and EagleEye [139] at
1.8× smaller peak memory.
IV. Tiny Training
In addition to inference, tiny on-device training is a
growing direction that allows us to adapt the pre-trained
model to newly collected sensory data after deployment.
By training and adapting locally on the edge, the model
can learn to continuously improve its predictions and
perform lifelong learning and user customization. By
bringing training closer to the sensors, it also helps to
protect user privacy when handling sensitive data (e.g.,
healthcare).
However, on-device training on tiny edge devices
is extremely challenging and fundamentally different
Figure 14. MCUNet has better visual wake word (VWW) accuracy versus peak SRAM trade-off (left). Compared to MCUNet [8],
MCUNet achieves better accuracy at 4.0× smaller peak memory. It achieves >90% accuracy under <32kB memory, facilitating
deployment on extremely small hardware. patch-based method expands the search space that can fit the MCU, allowing better
accuracy versus latency trade-off (right). (a) Trade-off: accuracy versus peak SRAM. (b) Trade-off: accuracy versus measured
latency.
Table 5.
MCUNet significantly improves Pascal VOC [134] object detection on MCU by allowing a higher input resolution.
Under STM32H743 MCU constraints, MCUNet-H7 improves the mAP by 16.9% compared to [8], achieving a record
performance on MCU. It can also scale down to cheaper MCU STM32F412 with only 256kB SRAM while still improving
mAP by 13.2% at 1.9× smaller peak SRAM and a similar computation.
MCU Model
H743 (∼$7)
F412 (∼$4)
24
Constraint
SRAM
<512kB
<256kB
IEEE CIRCUITS AND SYSTEMS MAGAZINE
Model
MbV2+CMSIS [8]
MCUNet [8]
MCUNet-H7
MCUNet-M4
#Param
0.87M
1.20M
0.67M
0.47M
MACs
34M
168M
343M
172M
peak SRAM
519kB
466kB
438kB
247kB
VOC
mAP
31.6%
51.4%
68.3%
64.6%
Gain
-
0%
+16.9%
+13.2%
THIRD QUARTER 2023
IEEE Circuits and Systems Magazine - Q3 2023

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q3 2023
