Disable test test that sporadically fails with OpenVINO on CI #29531
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [ ] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
dnn: add DynamicQuantizeLinear ONNX layer support #29018
### Pull Request Readiness Checklist
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and target platform(s)
- [x] There is an accuracy test
- [ ] There is a performance test
### Description
Implements the ONNX `DynamicQuantizeLinear` operator (opset 11) for the OpenCV DNN module.
**What it does:**
- Adds `QuantizeDynamicLayer` and `DequantizeDynamicLayer` layer classes
- Registers layers and adds ONNX importer dispatch for `DynamicQuantizeLinear`
- Computes scale and zero-point at runtime from activation min/max
- Quantizes FP32 input to int8 (stored as uint8 - 128, matching OpenCV convention)
**Framework limitation & workaround:**
Due to the single-dtype-per-layer constraint in `LayerData::dtype` (see #29017), the float32 scale output cannot be passed through the CV_8S blob pipeline directly. As a workaround, the scale is encoded as 4 raw bytes in a CV_8S `{1,4}` blob using `memcpy`, and decoded by the downstream `DequantizeDynamic` layer.
**Testing:**
- Custom accuracy tests reproduce all 3 ONNX conformance test cases: `test_dynamicquantizelinear`, `test_dynamicquantizelinear_max_adjusted`, `test_dynamicquantizelinear_min_adjusted`
- Each test verifies: quantized values, scale, zero point, and round-trip dequantize accuracy
- All existing quantization regression tests pass (42/42)
**Conformance tests:**
The 6 conformance tests for `DynamicQuantizeLinear` remain in the parser denylist (they were already denylisted before this PR) because the framework cannot produce mixed-type outputs. The custom tests provide equivalent coverage.
### Files changed
- `modules/dnn/include/opencv2/dnn/all_layers.hpp` — layer class declarations
- `modules/dnn/src/init.cpp` — layer registration
- `modules/dnn/src/onnx/onnx_importer.cpp` — ONNX import dispatch
- `modules/dnn/src/int8layers/quantization_utils.cpp` — layer implementations
- `modules/dnn/test/test_onnx_importer.cpp` — custom accuracy tests
bound tile axis and repeats length in onnx parseTile #29345
OpenCV Extra: https://github.com/opencv/opencv_extra/pull/1385
parseTile sizes repeats_vec from the input-0 rank, then fills it from fields of the model that are never checked against that size. In the tile-1 path the axis taken from the third input indexes repeats_vec directly, and in the tile>1 path the loop writes one entry per element of the repeats tensor. A crafted ONNX with an out-of-range axis, or a repeats tensor longer than the input rank, writes past repeats_vec while loading the model through readNetFromONNX.
The fix runs axis through normalize_axis, the same helper the squeeze and concat paths in this file already use, so a negative or oversized axis is rejected before the write, and it checks the repeats length equals the input rank before the loop. Keeping both bounds in the parser puts the check next to the write instead of trusting the model to be well formed. Before, a repeats tensor shorter than the rank was silently accepted; after, it is rejected, which matches the ONNX rule that repeats carries one entry per input dimension.
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
- [ ] The feature is well documented and sample code can be built with the project CMake
dnn: SIMD for transcendental activation layers (15 functors) 🧑💻🤖#2937
# dnn: SIMD for transcendental activation layers (15 functors) 🧑💻🤖
### Summary
Many `BaseDefaultFunctor`-based activation functors fall back to the scalar `BaseDefaultFunctor::apply`
— one `libm` call per element. This PR adds vectorized `apply()` overrides (universal intrinsics,
matching the existing `MishFunctor`/`GeluFunctor`/`SigmoidFunctor` pattern, each with a scalar tail) to
the **15 transcendental activations** that have a clean SIMD path:
| functor | uses | | functor | uses |
|---|---|---|---|---|
| TanH | `v_exp` (1−2/(e²ˣ+1)) | | Asinh | `v_log`,`v_sqrt` (sign·log(\|x\|+√(x²+1))) |
| Log | `v_log` | | Acosh | `v_log`,`v_sqrt` |
| Erf | `v_erf` | | Atanh | `v_log` |
| Exp | `v_exp` | | Softplus | `v_exp`,`v_log` |
| Sin | `v_sin` | | BNLL | `v_exp`,`v_log` (max(x,0)+log1p(e^−\|x\|)) |
| Cos | `v_cos` | | GeluApproximation | `v_exp` (tanh) |
| Tan | `v_sin`/`v_cos` | | Sinh / Cosh | `v_exp` |
Functors compile at the baseline SIMD width (SSE on x86, NEON on ARM — same as the existing SIMD
functors). `Sqrt`/`Floor`/`Ceil`/`Round`/`Abs`/etc. are intentionally **not** included — those are
single cheap ops the compiler already auto-vectorizes (measured `Sqrt` at 0.95×). `Asin`/`Acos`/`Atan`
are omitted (no `v_atan`/`v_asin`/`v_acos` intrinsic).
### Performance
Real `opencv_perf_dnn` `Layer_Activation` (added here), 8×256×128×100 = 26.2 M-elem CV_32F blob, A/B
vs the scalar fallback, median speedup:
| functor | M4 clang/NEON | A76 gcc/NEON | Threadripper gcc | Xeon W-2235 gcc | functor | M4 | A76 | TR | Xeon |
|---|---|---|---|---|---|---|---|---|---|
| TanH | 9.1× | 2.3× | 7.3× | 7.3× | Sinh | 3.3× | 3.2× | 5.2× | 5.0× |
| Cos | 8.8× | 3.0× | 3.3× | 3.0× | Acosh | 3.2× | 2.9× | 2.7× | 2.6× |
| Log | 7.8× | 3.0× | 2.5× | 2.4× | Cosh | 2.8× | 3.2× | 2.6× | 2.5× |
| Sin | 7.8× | 3.0× | 3.4× | 2.9× | Exp | 2.4× | 2.5× | 1.4× | 1.5× |
| GeluApprox | 7.6× | 2.4× | 6.6× | 5.8× | BNLL | 4.2× | 2.0× | 1.6× | 1.4× |
| Atanh | 7.0× | 3.6× | 6.0× | 5.2× | Softplus | 2.0× | 2.0× | 3.3× | 2.7× |
| Tan | 6.8× | 3.6× | 6.6× | 6.0× | Erf | 4.2× | 2.4× | 3.9× | 3.5× |
| Asinh | 3.9× | 2.4× | 5.8× | 5.7× | | | | | |
Every functor is a win on every tested out-of-order core (worst case 1.4×). On the **in-order
Cortex-A55** (a55-tuned build) all 15 measure 0.985–0.99× — neutral, within noise: the SIMD
polynomials are dependency-chain-bound, which an in-order pipeline can neither accelerate nor (here)
slow down. So: meaningful wins on out-of-order ARM + x86, no regression on in-order ARM.
### Accuracy
Uses the `v_exp`/`v_log`/`v_erf`/`v_sin`/`v_cos` polynomials OpenCV already ships and relies on (Gelu
via `v_erf`, Mish/Sigmoid via `v_exp`). Verified vs scalar `libm` on 320 K random elements per functor
with domain-correct inputs (Acosh x≥1, Atanh |x|<1, …): **max abs error ≤ 3e-5** (most ≤ 2e-6), zero
elements exceeding rel>1e-3 & abs>1e-4. ONNX conformance and the layer accuracy tests exercise these ops.
### Pull Request Readiness Checklist
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch (4.x)
- [ ] There is a reference to the original bug report and related work — N/A (perf improvement)
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable — perf test added (`Layer_Activation`, `SANITY_CHECK_NOTHING`); accuracy covered by existing ONNX conformance/layer tests
- [ ] The feature is well documented and sample code can be built with the project CMake — N/A (internal optimization, no new API)
guard the parsePadding NHWC swap on paddings.total()==8 and require the parseResize size tensor to hold height and width, matching the existing parseStridedSlice check
dnn: optimize blobFromImages using NEON intrinsics on ARM64 #29316
- This PR optimizes the HWC-to-NCHW conversion in blobFromImages() by adding NEON implementation.
- On x64, the MSVC compiler automatically vectorizes the existing scalar implementation, resulting in significantly better performance.
- On Windows ARM64, MSVC does not auto-vectorize this stride-3 access pattern, leading to a noticeable performance gap compared to x64.
- Added CV_NEON macro guard to ensure SIMD optimization is enabled on ARM64 architecture, leaving behavior on other platforms completely unchanged.
**Performance Improvements**
- This optimization significantly improves the performance of HWC-to-NCHW conversion on Windows ARM64.
<img width="823" height="254" alt="image" src="https://github.com/user-attachments/assets/84b22deb-22da-4607-8009-3405c2529b80" />
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [X] The PR is proposed to the proper branch
Reject initializers whose declared shape claims more elements than the tensor payload (raw_data or a typed *_data field) actually holds, before the Mat copy/convert reads them.
KleidiCV update to verison 26.03 #28744
KleidiCV release: https://gitlab.arm.com/kleidi/kleidicv/-/releases/26.03
Tuned DNN test threshold as resize linear produces slightly different result with the new KleidiCV version.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [ ] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
randomnormallike_layer.cpp registered itself via CV_DNN_REGISTER_LAYER_CLASS_STATIC at file scope. Nothing else in the translation unit was referenced from elsewhere, so when OpenCV is built statically (BUILD_SHARED_LIBS=OFF, as in the reporter's MSVC build) the linker drops the object file and the static-init registration never runs. The ONNX importer then sets layerParams.type = "RandomNormalLike" but getLayerInstance has no factory for that type, and parsing fails with "Can't create layer of type RandomNormalLike".
Switch to the standard pattern used by every other layer in the module: declare RandomNormalLikeLayer in all_layers.hpp, expose a static create() factory from the .cpp, and register it explicitly from initializeLayerFactory in init.cpp. Because init.cpp is referenced by the DNN module init path, this pulls in the layer .cpp regardless of build mode and the registration always runs.
The failure was incorrectly attributed to MSVC in the bug report. The bug is build-mode sensitive (static vs shared), not platform sensitive.
Verified locally on linux/gcc-13 by building modules/dnn and
opencv_test_dnn against the patched tree, then running opencv_test_dnn --gtest_filter='Test_ONNX_layers.RandomNormalLike_basic/0:Test_ONNX_layers.RandomNormalLike_complex/0'
Both tests pass; without the patch the same binary reproduces the
"Can't create layer of type RandomNormalLike" error from the report.
dnn: fix Resize initNgraph for two-input case #28724
## Summary
Fixes the issue #28707
When a Resize/Upsample layer has two inputs, the data tensor and a reference tensor whose **shape** defines the output spatial size, the OpenVINO/NGRAPH backend's `initNgraph()` was ignoring `nodes[1]` entirely and relying solely on the `outHeight`/`outWidth` member variables.
These variables are set by `finalize()` from the pre-computed output blob dimensions. However, when the output shape is determined dynamically at runtime from the second input, `finalize()` sets them from the live tensor, but the OpenVINO backend calls `initNgraph()` to build a static compiled graph. If the member variables are 0 at that point, the compiled `Interpolate` node gets hardcoded with `{0, 0}` output dimensions, causing CV_Assert failure: {N,C,0,0} vs {N,C,H2,W2}
DNN: Fix OpenVINO 2026 build failure due to ov::Tensor::data() const change #28592
Fixes: https://github.com/opencv/opencv/issues/28586
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
### PR Description
OpenVINO 2026 changed ov::Tensor::data() to return `const void*`
instead of `void*`. This causes a build error in
modules/dnn/src/op_inf_engine.cpp when constructing a cv::Mat
wrapper over Tensor memory.
Tensor memory for input/output blobs remains mutable, but the API
now enforces const-correct access. This patch casts away const with
an explicit comment to preserve existing zero-copy semantics and
restore compatibility with OpenVINO 2026.
Preserves existing zero-copy semantics of the OpenVINO backend without altering runtime behavior.
Tested by building OpenCV 4.x against OpenVINO 2026.0.0 on Ubuntu 24.04.
DNN: Fix Squeeze to remove all size-1 dims when axes is empty #28425Fixes#28424
OpenCV Extra: [opencv/opencv_extra#1308](https://github.com/opencv/opencv_extra/pull/1308)
This PR fixes the ONNX Squeeze operator to correctly remove all size-1 dimensions when `axes` is not provided, conforming to the ONNX specification.
### Details
Per [ONNX Squeeze specification](https://onnx.ai/onnx/operators/onnx__Squeeze.html):
> 'If axes is not provided, all the single dimensions will be removed from the shape.'
Previously, OpenCV DNN would not remove any dimensions in this case, causing shape mismatch errors with models like LaMa (inpainting).
### Example
```python
# Input: [1, 1, 2, 4]
# Squeeze with no axes attribute
# Before: [1, 1, 2, 4] ✗ (No change)
# After: [2, 4] ✓ (matches ONNX Runtime)
```
### Tests
Added `testONNXModels("squeeze_no_axes")` which validates this behavior with new test data.
opencv_extra_pr=opencv/opencv_extra#1324
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch (4.x for bug fixes)
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
- Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake
optimize(dnn): parallelize Slice layer implementation (4.x) #28511
### Summary
Backport of PR #28447 to 4.x branch.
### Description
This PR optimizes the SliceLayer implementation for strided inputs (where step > 1). The original implementation used a recursive element-wise copy (getSliceRecursive) for any strided slice, which was extremely inefficient.
This PR introduces:
- **Parallelization**: Uses `cv::parallel_for_` to parallelize the outermost dimension of the slice operation.
- **Memcpy Optimization**: Automatically detects "pseudo-contiguous" blocks in strided slices (e.g., slicing an outer dimension but keeping inner dimensions intact) and uses `std::memcpy` instead of scalar loops.
- **Refactoring**: Replaces the recursive function with a dedicated `ParallelSlice` loop body.
### Impact
Significant performance improvement for strided slice operations (common in detection heads, strided sampling, etc.).
### Benchmark Results
Tested on CPU with 20 threads.
| Test Case | Baseline (ms) | Optimized (ms) | Speedup |
| :--- | :--- | :--- | :--- |
| Strided Axis 0 [::2, ...] | 1.10 | 0.02 | **~55x** |
| Strided Axis 2 [..., ::2] | 1.15 | 0.11 | **~10.5x** |
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
- [ ] The feature is well documented and sample code can be built with the project CMake
dnn: improve robustness of MVN layer#28308
This change adds small defensive improvements to the MVN layer implementation:
->Guard against zero-sized OpenCL kernel launches
->Add input validation in finalize()
->Use int64 for FLOPS computation to avoid overflow
No functional or performance changes are intended.
### Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
- [ ] The feature is well documented and sample code can be built with the project CMake
The TopK layer was incorrectly rejecting K values equal to the input
dimension size. According to ONNX specification, K should be allowed
to equal the dimension size (to retrieve all elements).
Changed validation from 'K < input_shape[axis]' to 'K <= input_shape[axis]'
This fixes the error when loading YOLOv10 ONNX models that use TopK
with K equal to the dimension size.
Fixes#28445
docs: fix spelling errors in documentation and code #28301
- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [ ] There is a reference to the original bug report and related work
- [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
- [ ] The feature is well documented and sample code can be built with the project CMake
### Description
Fixed multiple spelling errors across documentation, comments, and code:
- 'colummn' → 'column' (cublas.hpp, 3 occurrences)
- 'points_per_colum' → 'points_per_column' (calib3d.hpp, 3 occurrences)
- 'Asignee' → 'Assignee' (sift files, 2 occurrences)
- 'compability' → 'compatibility' (face.hpp, 2 occurrences)
- 'orignal' → 'original' (aruco_detector.cpp)
- 'refrence' → 'reference' (chessboard.cpp)
- 'indeces' → 'indices' (stitching.hpp)
- 'OutputPrecison' → 'OutputPrecision' (test)
- 'tranform' → 'transform' (slice_layer.cpp, 3 occurrences)
Total: 24 fixes across 14 files. Documentation and comment changes only, no functional impact.
Alexander Smorkalov
Alexander Smorkalov
yyyisyyy11
Uwez Khan
Teddy-Yangjiale
uwezkhan
Tõnu Samuel
Pratham Kumar
Andrei Fedorov
Samaresh Kumar Singh
Dmitry Kurtaev
Kumataro
vrooomy
SamareshSingh
Weixie Cui
Prasad Ayush Kumar
Murat Raimbekov
ADITYA MISHRA
Hanbin Bae
Karnav Shah
Sikandar
Adrian Kretz
Vincent Rabaud