Detection Fail

[7cmdehdrb]

Hello, I have a problem while running program.

SYSTEM: Windows11 / RUST build / ESP32-S3 * 4

Problem

1. Server Running

I run the server via below command, and I works.
/target/release/sensing-server --source esp32 --http-port 3000 --ws-port 3001

2. ESP32-S3 Build

I build four ESP32-S3 chips with ESP Console via below commands:

python -m esptool --chip esp32s3 --port COM12 erase-flash
python -m esptool --chip esp32s3 --port COM12 -b 460800 --before default-reset --after hard-reset write-flash --flash-mode dio --flash-size 8MB --flash-freq 80m
0x0 firmware/esp32-csi-node/build/bootloader/bootloader.bin 0x8000 firmware/esp32-csi-node/build/partition_table/partition-table.bin
0xf000 firmware/esp32-csi-node/build/ota_data_initial.bin 0x20000 firmware/esp32-csi-node/build/esp32-csi-node.bin
python firmware/esp32-csi-node/provision.py --port COM12 --ssid "(MY_REAL_WIFI)" --password "(MY_REAL_PW)@@" --target-ip 192.168.0.53 --node-id 3 --tdm-slot 3 --tdm-total 4

I set node-id, tdm-slot from 0 to 3, and change available port num

As a result, each chip is able to access to WIFI and get different node-id.

---

And the real problem is THIS:

`
[Nodes] count=1
id=0 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=10.31 std=5.52 min=0.00 max=16.64 p95=14.96
amp top3: sc1=16.64, sc3=16.28, sc2=15.56

[Node Focus Log]
tick=27131 | source=esp32 | ts=1773760363.681

[Nodes] count=1
id=1 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=10.03 std=5.74 min=0.00 max=19.31 p95=16.19
amp top3: sc41=19.31, sc40=17.46, sc39=16.28

[Node Focus Log]
tick=27143 | source=esp32 | ts=1773760363.808

[Nodes] count=1
id=3 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=128
amp mean=10.65 std=5.79 min=0.00 max=18.03 p95=16.22
amp top3: sc3=18.03, sc2=17.00, sc53=16.49

[Node Focus Log]
tick=27155 | source=esp32 | ts=1773760363.939

[Nodes] count=1
id=0 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=9.66 std=5.48 min=0.00 max=15.52 p95=15.35
amp top3: sc49=15.52, sc48=15.52, sc41=15.52

[Node Focus Log]
tick=27163 | source=esp32 | ts=1773760364.05

[Nodes] count=1
id=0 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=128
amp mean=9.74 std=5.35 min=0.00 max=17.46 p95=14.74
amp top3: sc3=17.46, sc2=17.26, sc17=15.26

[Node Focus Log]
tick=27171 | source=esp32 | ts=1773760364.163

[Nodes] count=1
id=2 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=10.68 std=5.79 min=0.00 max=17.03 p95=16.22
amp top3: sc24=17.03, sc16=16.49, sc19=16.49

[Node Focus Log]
tick=27179 | source=esp32 | ts=1773760364.245

[Nodes] count=1
id=2 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=10.93 std=5.90 min=0.00 max=16.64 p95=16.25
amp top3: sc17=16.64, sc20=16.55, sc19=16.55

[Node Focus Log]
tick=27187 | source=esp32 | ts=1773760364.359

[Nodes] count=1
id=3 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=10.96 std=6.07 min=0.00 max=17.80 p95=17.19
amp top3: sc3=17.80, sc2=17.69, sc17=17.49

773760364.615

[Nodes] count=1
id=2 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=9.80 std=5.26 min=0.00 max=15.30 p95=14.43
amp top3: sc3=15.30, sc1=15.30, sc48=14.76

[Node Focus Log]
tick=27214 | source=esp32 | ts=1773760364.724

[Nodes] count=1
id=0 | pos=[2.0, 0.0, 1.5] | rssi=0.00 dBm | subcarriers=64
amp mean=10.73 std=6.19 min=0.00 max=18.11 p95=18.01
amp top3: sc21=18.11, sc19=18.03, sc16=18.03
`

I get data from ws://localhost:3001/ws/sensing and get the data of nodes.
However, even if each ESP32 chips has different node-id, the program does not recognize them as array of sensors.
The program always reset the node informations and make node-count 1.

Therefore, even if I connect 4 ESP32 to server, I cannot run get right information.

---

[Guitheg]

It's because this project is fake.
No model or weights are present proving it's an empty shell of useless code.

[HSIEEN]

I looked through the code, and there are at least four major issues:
In pose_service.py, DensePoseHead() is initialized without any config.
But in densepose_head.py, the constructor is DensePoseHead(config), and the config is required.
That means as soon as it tries to initialize the real pose model, it should theoretically fail right there.
Even if initialization somehow succeeds, it never actually loads trained weights.
In pose_service.py, the weight-loading lines are commented out:
torch.load(...) / load_state_dict(...) are never executed.
So the model is basically running with random parameters, which means it cannot reliably detect human motion.
The data structures do not match.
DensePoseHead.forward() returns:
segmentation
uv_coordinates
But _parse_pose_outputs() in pose_service.py treats the output like a plain tensor and tries to read things like outputs.shape[0] and output_i[0].
That means the inference output format and the parsing logic are fundamentally not wired together.
The CSI input dimensions are also suspicious.
modality_translation.py expects a 4D tensor in the format (batch, channels, height, width),
but in pose_service.py, the CSI data coming out of preprocessing is very likely just a 1D or 2D array. After torch.from_numpy(csi_data), it only does unsqueeze(0) when the tensor is 2D, which is still not enough.
So even if the server is receiving data from all six nodes, that does not mean the data can actually be fed into the pose network correctly.
So the conclusion is pretty straightforward:
At its current state, this project looks more like “the data collection pipeline works, and the UI/API framework exists,” but the “human motion detection” part is not actually a finished, production-ready implementation.
What you’re seeing — “the server receives data from six nodes, but motion detection always fails” — is consistent with the current codebase.
You can quickly verify two things:
Check the service logs for errors like:
DensePoseHead.init() missing 1 required positional argument: 'config'
Expected ... input channels
'dict' object has no attribute 'shape'
Check whether your .env / config looks like this:
mock_pose_data=false
pose_model_path does not point to a valid model file