TinySense+ Code
WI-FI SENSING · CSI COMPRESSION · EDGE INTELLIGENCE

TinySense / TinySense+

Scalable Wi-Fi sensing,
from fewer bits.

Preserve the information that matters for human pose estimation.
Compress CSI at the edge. Reconstruct and sense at the server.

Read TinySense+ TinySense · PerCom 2026 Explore code
8.00 → 3.21bits per index with rate-aware training
44.01%PCK20 at 3.21 bits per index
5,840×compression at 37.92% PCK20
HUMAN POSE ESTIMATION FROM WI-FI Reference Wi-Fi prediction
MM-Fi examples comparing reference pose keypoints in red and Wi-Fi predictions in blue at compression ratios 4, 84, and 570.
Pose estimates from compressed CSI on MM-Fi. The shown examples illustrate increasing compression; quantitative results are reported below.
01 / PUBLICATIONS

Two papers. A shared foundation.

TinySense introduces the sensing representation.
TinySense+ extends how it is trained and transmitted.

CONFERENCE PAPERIEEE PerCom 2026

TinySense: Effective CSI Compression for Scalable and Accurate Wi-Fi Sensing

Toan Gian, Dung T. Tran, Viet Quoc Pham, Francesco Restuccia, and Van-Dinh Nguyen

Task-aware VQGAN compression with K-means codebook resizing and Transformer-based recovery of missing indices.

Paper ↗arXiv ↗Code ↗Cite ↓
EXTENDED PAPERPreprint · 2026

TinySense+: Entropy-Aware CSI Compression for Scalable and Adaptive Wi-Fi Sensing

Phuc H. Nguyen, Toan Gian, Cuong D. Do, Symeon Chatzinotas, and Van-Dinh Nguyen

Adds a learned entropy model, rate-aware training, arithmetic coding, and adjustable prefix transmission without encoder retraining.

Paper ↗alphaXiv ↗Code ↗Cite ↓
THE RESEARCH QUESTION

How much CSI do we
need to send?

Wi-Fi human pose estimation uses subtle changes in channel state information to infer body motion. Continuously sending high-dimensional CSI to a server can make communication the bottleneck.

TinySense learns a discrete representation that preserves pose information. TinySense+ also optimizes how predictable those indices are, reducing coded length and allowing a device to send a shorter prefix when the communication budget changes.

Follow the pipeline
02 / METHOD

Learn. Compress. Sense.

A lightweight encoder at the edge.
A shared codebook and pose estimator at the server.

Wi-Fi transmitter and receiver acquire CSI; local devices encode latent indices and send them to an edge server for human pose estimation.
Edge–server Wi-Fi sensing: transmit a learned latent representation instead of raw CSI.
01

Task-aware representation

The encoder maps CSI to a VQ codebook. Reconstruction and keypoint supervision retain information useful for pose estimation.

TinySense + TinySense+
02

Entropy-aware transmission

A causal Transformer models index probabilities. Arithmetic coding and a differentiable rate objective reduce expected code length.

Added in TinySense+
03

Adjustable context

Send a selected prefix of the latent sequence. At the server, prefix-conditioned latent completion infers the omitted suffix.

Added in TinySense+
Inside TinySense+ The entropy model and prefix-completion pipeline
TinySense+ architecture showing vector quantization, a rate objective, arithmetic encoder and decoder, and autoregressive completion of an omitted suffix.
AE / AD: arithmetic encoder / decoder. Solid arrows denote data flow; dashed arrows denote learning and modeling flow.
What changes from TinySense to TinySense+?
ComponentTinySenseTinySense+
TrainingTask-aware VQGANAdds rate regularization
Index codingFixed-length binaryCausal entropy model + arithmetic coding
Rate controlK-means codebook resizingAdds adjustable prefix length
RecoveryPredicts erased indicesCompletes the deliberately omitted suffix
03 / REPORTED RESULTS

Fewer bits. Useful pose information.

Component ablation on MM-Fi.
Values reproduced from the TinySense+ paper.

Transmission cost bits / index ↓
TinySense
8.00
+ entropy coding
4.82
+ rate-aware training
3.21

Entropy coding preserves the index sequence. Rate-aware training changes the learned representation.

RATE-AWARE TRAINING59.9%

fewer bits per index

Relative to the 8.00-bit baseline in this ablation, with PCK20 increasing from 41.21% to 44.01%.

Training and transmission ablation · MM-Fi
ConfigurationBits / index ↓Compression η ↑PCK20 (%) ↑MPJPE ↓
TinySense8.0021541.21170.98
+ entropy model4.8235641.21170.98
+ rate-aware training3.2153544.01170.17
+ prefix completion (75%)2.4271042.73173.91
+ prefix completion (50%)1.691,01540.36178.42
EXPLORE THE REPORTED OPERATING POINTS

Send a shorter prefix.

The encoder stays fixed. The server completes the omitted suffix. Select one of three measured settings from the ablation.

Transmitted prefix
Transmitted prefix Completed suffix
3.21bits / index
535×compression
44.01%PCK20

Reported values · MM-Fi · no interpolation

η is the compression ratio relative to the paper’s raw-CSI representation. These ablation settings are distinct from the 5,840× operating point highlighted above. Evaluation uses MM-Fi and WiPose; unseen-site generalization without adaptation remains unverified.

04 / FROM MODEL TO SYSTEM

An edge–server prototype.

Raspberry Pi and Jetson Nano devices.
An RTX 8000 server.

Experimental setup with Raspberry Pi and Jetson Nano edge devices connected to a workstation with an RTX 8000 GPU.
Hardware prototype used for latency, local energy, and networking measurements.

Measure the whole sensing pipeline.

The experiments evaluate communication cost alongside pose accuracy and device computation. Entropy coding adds computation while reducing payload; the practical benefit depends on the operating point and hardware.

Read the evaluation in the paper ↗
05 / BUILD ON THIS WORK

Citation & resources.

View repository ↗
@misc{nguyen2026tinysenseplus,
  title = {TinySense+: Entropy-Aware CSI Compression for
           Scalable and Adaptive Wi-Fi Sensing},
  author = {Nguyen, Phuc H. and Gian, Toan and Do, Cuong D.
            and Chatzinotas, Symeon and Nguyen, Van-Dinh},
  year = {2026},
  note = {Preprint},
  url = {https://www.alphaxiv.org/abs/2610.tinysense-entropy-aware-wifi-sensing}
}

TinySense+ is listed as a preprint. Use the citation corresponding to the version you build on.