WiFi DensePose: What the CSI experiments actually establish

WiFi DensePose maps measured radio-channel data to dense human pose estimates in a research setup. Its evidence concerns a specified antenna arrangement, channel measurements, training supervision, and evaluation split, rather than an automatic capability of every home router.

The original DensePose vision work maps human pixels in an RGB image to positions on a body surface. That is a richer target than detecting a person or identifying a few joints. A pose representation is still different from a photograph of a person’s clothes or an independently verified identity. (DensePose: Dense Human Pose Estimation In The Wild - arXiv:1802.00434)

WiFi DensePose: What the CSI experiments actually establish

DensePose From WiFi learns a mapping from signal phase and amplitude to body-region and surface coordinates. Its setup describes three transmitters and three aligned receivers. The paper notes that a three-antenna router configuration could supply that arrangement, but the sensing pipeline also needs the appropriate channel measurements and a trained model. An antenna count alone is insufficient implementation evidence. (DensePose From WiFi study)

Channel State Information describes a radio channel’s measured response across the links and frequencies made available by the hardware. Here, amplitude and phase are model inputs; an ordinary connection-strength indicator is not an equivalent input tensor. The generic term has a background explanation in Channel State Information - Wikipedia, while the study defines the measurements used by its model. (DensePose From WiFi study)

The multiple-antenna arrangement matters because it supplies measurements across antenna pairs. MIMO - Wikipedia explains the broader communications term; the research setup and its data tensor remain the relevant evidence for this sensing application. The earlier Person-in-WiFi study likewise used transmitter and receiver sets with three antennas each, with image-derived labels for segmentation and pose estimation. (Person-in-WiFi study)

Separate camera-free inference from training supervision

The WiFi studies use image-based information during training to teach a model the target pose or segmentation. Their trained inference path can use radio measurements, but that does not mean the training labels or environment preparation disappear from the system’s provenance. The transfer of a Computer Vision - Wikipedia task to radio input is the research contribution, rather than proof of a plug-in replacement for every camera deployment. (DensePose From WiFi study; Person-in-WiFi study)

The DensePose study describes a split that trains on 15 spatial layouts and tests on one unseen layout. That supplies evidence about the tested split. It does not establish performance in every furnished home, with every subject, antenna placement, construction material, and interference condition. Keep same-environment results distinct from held-out-environment results when evaluating whether a model can be reused. (DensePose From WiFi evaluation)

Indoor signals combine paths through the environment. Multipath Propagation - Wikipedia explains that background concept. The resulting channel measurements reflect more than a person’s pose, so the evaluation needs to test how the learned model handles scene changes. This is a practical inference from the studies’ layout-dependent inputs and evaluation designs, not an additional reported accuracy figure.

Through-wall imaging is another experiment

The separate Through-Wall Imaging based on WiFi Channel State Information study evaluates an image-generation system using ESP32-S3 development boards and directional antennas. Its result should be attached to that equipment and protocol. It is not a direct performance measurement of the DensePose model or a general consumer mesh network. (Through-wall CSI imaging study)

That paper also warns that identities in generated images are hallucinated from training data. A visually convincing output therefore cannot be treated as a camera observation identifying the person behind a barrier. The image-generation target, pose-estimation target, and evidence needed for identity attribution are separate. (Through-wall CSI imaging study)

The DensePose paper discusses radio alternatives alongside cameras, LiDAR, and radar. Radar - Wikipedia supplies background terminology; a comparison of experimental apparatus does not establish universal through-wall range, lower cost, or undetectable monitoring. (DensePose From WiFi study)

Check the channel-data path before the model

Nexmon CSI is a concrete extraction project. Its documentation lists supported Broadcom chips, firmware versions, and devices, including particular phones, Raspberry Pi configurations, and an Asus router. It describes per-frame CSI extraction and device-specific setup. The existence of that support matrix does not make every router compatible, and successful extraction does not by itself implement DensePose. (Nexmon CSI support and extraction documentation)

For a proposed experiment, obtain evidence for each layer:

These are proposed review checks derived from the documented systems. They are not a reproduction performed for this article.

A useful demonstration can use inexpensive communications hardware while still depending on a carefully assembled sensing system. The evidence supports investigating WiFi-based pose and imaging under those conditions. It does not support a universal accuracy table, a claim that existing home networks need only a firmware switch, or a conclusion that walls everywhere have ceased to protect visual privacy.