infer and preproc a grayscale model by Rtms1984 · Pull Request #379 · TexasInstruments/edgeai-gst-plugins

Rtms1984 · 2026-06-03T14:55:24Z

Add grayscale model preprocessing and inference support

Rtms1984 · 2026-06-04T11:30:02Z

Across the three elements, no new user-facing properties were added — the changes are all internal plumbing to support a new GRAY8 / single-channel path (preproc elements) and to handle 4-D model outputs more correctly (inferer).

tiovxdlpreproc

Sink caps: adds GRAY8 to TIOVX_DL_PRE_PROC_SUPPORTED_FORMATS_SINK
New struct field gboolean gray8_cpu_bypass
init_module: for GRAY8 input, output tensor channel count is forced to 1 and the bypass flag is enabled
New transform override that, in bypass mode, does CPU (pixel-mean)*scale → float32, copies timestamps, and skips the TIOVX kernel (Could not add grayscale support since the code is closed source).
No new GObject properties. Existing scale-* / mean-* are reused; only scale[0]/mean[0] matter in the GRAY8 path.

tidlpreproc

Sink caps: {RGB, NV12} → {RGB, NV12, GRAY8}
New struct field gint num_channels on _GstTIDLPreProc, default 3
parse_model now also extracts the channel count from the model's input tensor shape and bumps the NCHW shape size check from >=3 to >=4 and NHWC from >=2 to >=3.
New GRAY8 branch in transform: per-pixel (v - mean[0]) * scale[0] with a switch over tensor_data_type writing int8/uint8/int16/uint16/int32/uint32/float32 directly into the output tensor, honoring in_stride_y
No new GObject properties. Existing props (model, scale-, mean-, channel-order, data-type, tensor-format, out-pool-size) are unchanged.

tidlinferer

New struct field guint input_channels on _GstTIDLInferer, populated in transform_caps for both NHWC/NCHW layouts. It's tracked internally but not yet exposed as a property or in caps.
Set_output_params reworked:

Adds a special-case for dim == 4 outputs: current_height = shape[2], current_width = shape[1] * shape[3]. The previous "first two non-1 dims" heuristic is kept only as the fallback for non-4D outputs.
transform_caps fixation now also explicitly fixates num-dims to NUM_TENSOR_DIMS on both src and sink directions.
No new GObject properties. Existing: model, in-pool-size, out-pool-size, debug-level, plus target when ENABLE_TIDL.

cshilwant · 2026-06-11T16:02:53Z

@Rtms1984

Thanks for the pull request! I have a couple of minor comments before we can go ahead and integrate the change:

Could you please add a Signed-off-by tag to your commit message? You can easily do this by amending your commit using git commit -s --amend and force-pushing.
Update the destination branch: Please change the destination branch of this PR from main to develop.
Can we have some tests logs around how GRAY8 FORMAT was validated for preproc.

Once these are updated, we'll get this merged right away.

Thanks again for your contribution!

Signed-off-by: talebso <sogand.talebi@sick.se>

Rtms1984 · 2026-06-15T09:14:33Z

Here is the logs from my tests on a grayscale MobilenetV2 classification model with both tidlpreproc and tiovxdlpreproc.
gst-logs.txt

reesegrimsley

PR approved after test-case details provided

sinha-shreyash

@Rtms1984 , I have few feedback points after having a look at your PR,

Can you please write a appropriate commit msg which describes the changes you are making in more detail.
Please follow the code convention present across entire repo, changing it for few files does not makes sense to me
I see lot of the crux changes are present in transform code, instead it should be handled in utils under edgeai-apps-utils (just like NV12/RGB image) to have uniformity across the code so that we can find all the preproc related operations under one page.

sinha-shreyash · 2026-07-01T07:08:17Z

 #include "gst-libs/gst/tiovx/gsttiovxutils.h"

-#endif //ENABLE_TIDL
+#endif                          //ENABLE_TIDL


trailing white spaces

sinha-shreyash · 2026-07-01T07:08:30Z

 #define ARM_MAX_OUTPUT_WIDTH  1001
 #define ARM_MAX_OUTPUT_HEIGHT 1001
-#endif // NOT ENABLE_TIDL
+#endif                          // NOT ENABLE_TIDL


trailing white spaces

sinha-shreyash · 2026-07-01T07:08:45Z

  gsize
      output_size;
-#endif //ENABLE_TIDL
+#endif                          //ENABLE_TIDL


trailing white spaces

sinha-shreyash · 2026-07-01T07:08:51Z

 #ifdef ENABLE_TIDL
  PROP_TARGET,
-#endif //ENABLE_TIDL
+#endif                          //ENABLE_TIDL


trailing white spaces

sinha-shreyash · 2026-07-01T07:09:32Z

 gst_ti_dl_inferer_get_property (GObject * object, guint prop_id,
    GValue * value, GParamSpec * pspec);

-static guint


can we please maintain the code format we are following across the repository, simply changing it for this file does not makes sense to me

sinha-shreyash · 2026-07-02T06:06:15Z

Can we please follow the code format followed across entire repo

sinha-shreyash · 2026-07-02T07:48:21Z

-          &&
-          self->pre_proc_config->inputTensorShapes[0].size() >= 3) {
+      if (self->pre_proc_config->inputTensorShapes.size () > 0
+          && self->pre_proc_config->inputTensorShapes[0].size () >= 4) {


why is this changed to 4 ? should not this be 3 same as NHWC format

sinha-shreyash · 2026-07-02T08:15:31Z

  } else if (GST_VIDEO_FORMAT_RGB == GST_VIDEO_FRAME_FORMAT (&in_frame)) {
    dlPreProcess_RGB_image (self->pre_proc_image_params);
+  } else if (GST_VIDEO_FORMAT_GRAY8 == GST_VIDEO_FRAME_FORMAT (&in_frame)) {
+    /* Single-channel CPU normalization: (pixel - mean) * scale.


why not add dlPreProcess_GRAY_image function in edgeai-apps-utils ?

sinha-shreyash · 2026-07-02T08:16:54Z

please follow the same code format used across repositories

sinha-shreyash · 2026-07-02T09:32:01Z

  return TRUE;
 }

+static


I think all this process is handled internally in preproc utils defined in edgeai-apps-utils, should not it be added there instead ?

@cshilwant, please comment more on this part

Rtms1984 force-pushed the grayscale-model-inference branch from 8e6aeac to 5018e33 Compare June 12, 2026 08:14

Rtms1984 changed the base branch from main to develop June 12, 2026 08:15

infer and preproc a grayscale model

0b63733

Signed-off-by: talebso <sogand.talebi@sick.se>

Rtms1984 force-pushed the grayscale-model-inference branch from 5018e33 to 0b63733 Compare June 15, 2026 05:11

cshilwant approved these changes Jun 25, 2026

View reviewed changes

cshilwant requested review from reesegrimsley and sinha-shreyash June 25, 2026 12:03

reesegrimsley approved these changes Jun 25, 2026

View reviewed changes

sinha-shreyash requested changes Jul 2, 2026

View reviewed changes

Uh oh!

Conversation

Rtms1984 commented Jun 3, 2026

Uh oh!

Rtms1984 commented Jun 4, 2026

Uh oh!

cshilwant commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Rtms1984 commented Jun 15, 2026

Uh oh!

reesegrimsley left a comment

Choose a reason for hiding this comment

Uh oh!

sinha-shreyash left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

cshilwant commented Jun 11, 2026 •

edited

Loading