Skip to content

Document ultra light face detector without postproc variant #14

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ocrete wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Document ultra light face detector without postproc variant #14

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
232adb9
Add tensor for Ultra-Light-Fast-Generic-Face-Detector without post pr…
ocrete Nov 20, 2025
581b090
Add tensor-groups for Ultra-Light-Fast-Generic-Face-Detector
ocrete Nov 20, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
52 changes: 52 additions & 0 deletions tensor-groups/ultra-lightweight-face-detection-rfb-320-v1-without-postproc-out.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
# Classification
- tensor-group: yes
- layer-type: output
- use-case: face-detection

# Description

A specialized model that detects faces

Ultra-Light-Fast-Generic-Face-Detector Output Tensors:

| Name |shape |
|--- |--- |
| [boxes] |1 x COUNT x 4 |
| [scores] |1 x COUNT x 2 |

Where COUNT is a value selected at training time.

# Tensor Decoding Logic

```
Foreach i in count:
X = boxes_processed(i, 0)
Y = boxes_processed(i, 1)
W = boxes_processed(i, 2)
H = boxes_processed(i, 3)
S = scores[i][1]
If S > threshold:
detection_candidates[j++] = [X, Y, W, H, S]
detections = non_max_suppression(detection_candidates)

```

Where X, Y, W and H are values between 0 and 1. The boxes have been processed
from the anchors.

# External References
* https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB

# Models
* [Model source](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master)
* [ONNX pre-trained model](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/models/)

# Tensor Decoders
|Framework | Links |
|--- |--- |
|GStreamer | [perm](https://gitlab.freedesktop.org/gstreamer/gstreamer/-/blob/main/subprojects/gst-plugins-bad/gst/tensordecoders/gstfacedetectortensordecoder.c) |


[boxes]: /tensors/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-boxes-without-postproc.md
[scores]: /tensors/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-scores.md

2 changes: 2 additions & 0 deletions tensor-id-register.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
|classification-generic-out | [details](/tensors/classification-generic-out.md) |
|classification-generic-softmaxed-out | [details](/tensors/classification-generic-softmaxed-out.md) |
|ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-scores | [details](/tensors/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-scores.md) |
|ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-boxes-without-postproc | [details](/tensors/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-boxes-without-postproc.md) |
|yolo-v8-pose-out | [details](/tensors/yolo-v8-pose-out.md) |
|yolo-v8-segmentation-out-detections | [details](/tensors/yolo-v8-segmentation-out-detections.md) |
|yolo-v8-segmentation-out-protos | [details](/tensors/yolo-v8-segmentation-out-protos.md) |
Expand All @@ -19,4 +20,5 @@
|--- |--- |
|ssd-mobilenet-v1-variant-1-out | [details](/tensor-groups/ssd-mobilenet-v1-variant-1-out.md) |
|ultra-lightweight-face-detection-rfb-320-v1-variant-1-out | [details](/tensor-groups/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out.md) |
|ultra-lightweight-face-detection-rfb-320-v1-without-postproc-out | [details](/tensor-groups/ultra-lightweight-face-detection-rfb-320-v1-without-postproc-out.md) |
|yolo-v8-segmentation-out | [details](/tensor-groups/yolo-v8-segmentation-out.md) |
2 changes: 2 additions & 0 deletions tensors/ssd-mobilenet-v1-variant-1-out-boxes.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,8 @@
- use-case: object-detection
- part-of-tensor-groups:
- [ssd-mobilenet-v1-variant-1-out](/tensor-groups/ssd-mobilenet-v1-variant-1-out.md)
- [ultra-lightweight-face-detection-rfb-320-v1-variant-1-out](/tensor-groups/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out.md)


# Description
Location of objects detected
Expand Down
80 changes: 80 additions & 0 deletions ...a-lightweight-face-detection-rfb-320-v1-variant-1-out-boxes-without-postproc.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,80 @@
# Classification

- tensor-group: no
- layer-type: output
- use-case: face-detection
- part-of-tensor-groups:
- [ultra-lightweight-face-detection-rfb-320-v1-without-postproc-out](/tensor-groups/ultra-lightweight-face-detection-rfb-320-v1-without-postproc-out.md)

# Description
Location of faces detected

## Boxes Tensor

- tensor-shape: 1 x COUNT x 4
- tensor-datatype: float32
- tensor-id: ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-boxes-without-postproc

### Encoding

Scheme: (top left X, top left Y, width, height)

The COUNT is variable and depends on the model training.

Other constants are extracted from the training process:
- CENTER_VARIANCE = 0.1
- SIZE_VARIANCE = 0.2
- COUNT

The COUNT can be retrieved from the tensor dimensions.

Each entry in the table is the variance on the matching anchor which
is also defined at training time. Each anchor is defined as the
following tuple:
`(anchor-center-x, anchor-center-y, anchor-width anchor-height)`

Once the following formulas have been applied, the bounding boxes
center is calculated along with its size.

```
bounding-box-center-x = top-left-x * CENTER_VARIANCE * anchor-width + anchor-center-x
bounding-box-center-y = top-left-y * CENTER_VARIANCE * anchor-height + anchor-center-y
Comment on lines +40 to +41
Copy link
Collaborator

@dmorin1 dmorin1 Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't really understand this formula but maybe once you clarify anchor-center-x and anchor-center-y. It will be more clear.

Copy link
Collaborator Author

@ocrete ocrete Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The formula is taken from the tensor decoder... It moves the anchor box by a certain value I think

bounding-box-width = expf (width * SIZE_VARIANCE) * anchor-width
Copy link
Collaborator

@dmorin1 dmorin1 Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

what is expf ?

Copy link
Collaborator Author

@ocrete ocrete Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

it's expf(x) is the C function that does e^(x) .. would that be clearer ?

Copy link
Collaborator Author

@ocrete ocrete Nov 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

something like:

Suggested change
bounding-box-width = expf (width * SIZE_VARIANCE) * anchor-width
bounding-box-width = e^(width * SIZE_VARIANCE) * anchor-width

bounding-box-height = expf (height * SIZE_VARIANCE) * anchor-height
```

|box-1 | box-1 | box-1 | box-1 | box-2 | box-2 | box-2 | box-2 | ... | COUNT|COUNT|COUNT|COUNT|
|--- |--- |--- |--- |--- |--- |--- |--- |--- |--- |--- |--- |--- |
| top-left-X | top-left-Y | width | height | top-left-X | top-left-Y | width | height | ... top-left-X | top-left-Y | width | height |

Memory layout of tensor data:

|Index |Value |
|--- |--- |
| - | - |
|0 |top-left-x |
|1 |top-left-y |
|2 |width |
|3 |height |
|4 |top-left-x |
|5 |top-left-y |
|6 |width |
|7 |height |
|... | ... |
|(COUNT - 1) x 4 |top-left-x |
|(COUNT - 1) x 4 + 1 |top-left-y |
|(COUNT - 1) x 4 + 2 |width |
|(COUNT - 1) x 4 + 3 |height |



# Models

* [Model source](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master)
* [ONNX pre-trained model](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/models/)

# Tensor Decoders
|Framework | Links |
|--- |--- |
|GStreamer | [perm](https://gitlab.freedesktop.org/gstreamer/gstreamer/-/blob/main/subprojects/gst-plugins-bad/gst/tensordecoders/gstfacedetectortensordecoder.c) |

1 change: 1 addition & 0 deletions tensors/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out-scores.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
- use-case: face-detection
- part-of-tensor-groups:
- [ultra-lightweight-face-detection-rfb-320-v1-variant-1-out](/tensor-groups/ultra-lightweight-face-detection-rfb-320-v1-variant-1-out.md)
- [ultra-lightweight-face-detection-rfb-320-v1-without-postproc-out](/tensor-groups/ultra-lightweight-face-detection-rfb-320-v1-without-postproc-out.md)

# Description

Expand Down