Skip to content

HIGH_IMPACT_PLAN.md

Latest commit

 

History

History
183 lines (134 loc) · 4.71 KB

HIGH_IMPACT_PLAN.md

File metadata and controls

183 lines (134 loc) · 4.71 KB

High-Impact Conformance Plan

Analysis: 15,277 Tests Across 69 Categories

Top 15 Test Groups by Count

Rank Group Tests Current Status Impact
1 Convolve 1009 Unknown 🔴 Critical
2 Scale 982 Partial 🔴 Critical
3 HarrisCorners 433 Unknown 🟡 High
4 Remap 380 Unknown 🟡 High
5 WarpPerspective 361 Unknown 🟡 High
6 WarpAffine 305 Partial 🟡 High
7 Image 230 Many failing 🟡 High
8 NonLinearFilter 172 Unknown 🟢 Medium
9 vxMapImagePatch 156 Unknown 🟢 Medium
10 Graph 118 Working ✅ 🟢 Medium
11 vxCopyImagePatch 117 Unknown 🟢 Medium
12 WeightedAverage 102 Unknown 🟢 Medium
13 Scalar 102 Working ✅ 🟢 Medium
14 vxAddSub 76 Unknown 🟢 Low
15 ColorConvert 56 Failing 🟢 Low

Top 5 groups = 3,485 tests (23% of total)

Strategic Fix Plan

Phase 1: Fix Image Infrastructure (Unblocks Groups 1-7)

Estimated Impact: +2,000 tests

Problem: Many vision tests fail due to:

  1. Virtual image handling
  2. Image format support (S16 needed for Sobel)
  3. Image patch mapping

Fixes:

  1. ✅ Fix virtual image creation/validation
  2. ✅ Add S16 format support for gradients
  3. ✅ Fix image patch mapping functions
  4. ✅ Handle image from ROI

Files:

  • openvx-image/src/c_api.rs
  • openvx-core/src/image_format.rs

Phase 2: Geometric Transforms (Groups 2, 4, 5, 6)

Estimated Impact: +2,028 tests

Groups: Scale (982), Remap (380), WarpPerspective (361), WarpAffine (305)

Current Status: Partial - some implementations exist but accuracy issues

Fixes:

  1. ✅ Fix bilinear interpolation accuracy
  2. ✅ Fix border handling in geometric transforms
  3. ✅ Ensure output format matches input

Files:

  • openvx-vision/src/geometric.rs
  • openvx-vision/src/transform.rs

Phase 3: Convolution (Group 1 - 1,009 tests)

Estimated Impact: +1,009 tests

Current Status: Custom convolution likely has kernel handling issues

Fixes:

  1. ✅ Fix vxConvolve kernel size handling
  2. ✅ Ensure proper normalization
  3. ✅ Fix convolution data type handling

Files:

  • openvx-vision/src/filter_simd.rs
  • openvx-core/src/c_api_data.rs (convolution objects)

Phase 4: Feature Detection (Group 3 - HarrisCorners)

Estimated Impact: +433 tests

Current Status: Unknown, but algorithms exist in code

Fixes:

  1. ✅ Fix HarrisCorners parameter handling
  2. ✅ Ensure gradient computation accuracy
  3. ✅ Fix corner response calculation

Files:

  • openvx-vision/src/feature_detection.rs
  • openvx-vision/src/object_detection.rs

Phase 5: Quick Wins (Groups 8-15)

Estimated Impact: +800 tests

Quick fixes for:

  • NonLinearFilter (172) - median filter
  • vxMapImagePatch (156) - likely patch access
  • vxCopyImagePatch (117) - data copying
  • ColorConvert (56) - fix YUV/RGB conversions

Files:

  • Various vision kernel files

Total Potential Impact

Phase Tests Priority
Phase 1: Image Infrastructure +2,000 🔴 Critical
Phase 2: Geometric Transforms +2,028 🔴 Critical
Phase 3: Convolution +1,009 🟡 High
Phase 4: Feature Detection +433 🟡 High
Phase 5: Quick Wins +800 🟢 Medium

Total: +6,270 tests (41% of total suite)

Immediate Action Plan

Step 1: Image Format Support (This Session)

Task: Add S16 format for Sobel/gradient outputs Agent: image-format-agent Time: 30 min Impact: Unlocks 306+ tests (Sobel3x3, gradients)

Step 2: Virtual Image Fix (This Session)

Task: Fix vxCreateVirtualImage validation Agent: virtual-image-agent Time: 30 min Impact: Unlocks 230 Image tests

Step 3: Geometric Transform Accuracy (Next Session)

Task: Fix Scale, WarpAffine, WarpPerspective Agent: geometric-agent Time: 1 hour Impact: +1,648 tests

Step 4: Convolution Kernel (Next Session)

Task: Fix vxConvolve implementation Agent: convolution-agent Time: 1 hour Impact: +1,009 tests

Current Working Foundation

✅ Already Working:

  • GraphBase: 14/14 (100%)
  • Box3x3: 23/23 (100%)
  • Gaussian3x3: 9/9 (100%)
  • Core reference counting
  • Kernel loading/unloading

These prove the infrastructure is solid!

Risk Assessment

High Risk:

  • Image format changes (may break existing tests)
  • Convolution accuracy (complex math)

Low Risk:

  • Virtual image fixes (isolated)
  • Border handling tweaks

Mitigation:

  • Run tests after each fix
  • Commit frequently
  • Use feature flags if needed