# 09 — Stage 6 BEV cleanup Date: 16 July 2026 Project: `classical-lidar-lane-detection` Plan: [`07-Plan-From-Here-On.md`](07-Plan-From-Here-On.md) ## Executive summary Here is the short version of what I built and what the result actually means. - **The setup:** Stage 6 now takes the canonical path we agreed on: current Stage 3, raw Stage-4 reflectivity, and raw Stage-5 global Otsu. It does not consume the Stage-5.1 top-decile/row-budget proposal. I evaluated it on the same 40-frame development slice used in the earlier audits: ten evenly spaced labeled frames from each of four sequences. Thirty-nine were readable; the same known malformed sequence-4 PCD was recorded and skipped. - **The problem:** Raw Stage 5 contains useful lane returns mixed with broad scan structures and scattered bright road clutter. The first, geometry-only cleanup was much too aggressive because genuine lane paint often appears in one LiDAR frame as short transverse scan fragments—not as neat little forward-facing lines. Rejecting every locally transverse component destroyed recall. I discarded that configuration instead of polishing a bad result. - **What I implemented:** Candidates are rasterized at `288 x 288`, which is `0.16 m` forward by `0.08 m` lateral per cell. Stage 6 closes only short gaps in the forward direction, extracts connected components, records their physical span, orientation, elongation, support density, reflectivity, and point support, then conservatively removes weak long/wide structures. A separate chain pass can restore a weak compact fragment only when multiple narrow fragments line up ahead/behind it. The final mask is also reduced to K-Lane's `144 x 144` grid, while the `288 x 288` mask remains available for Stage 7. - **The outcome:** On this development slice, K-Lane-tolerant precision rises from **22.59% to 25.75%**, recall changes from **13.84% to 13.63%**, and F1 rises from **17.16% to 17.83%**. Put another way, cleanup removes 1,777 tolerant false-positive cells while losing 46 true-positive label hits. It keeps 98.48% of the baseline true positives. Twenty-nine frames improve, five are unchanged, and five get worse. - **The honest interpretation:** Stage 6 is doing the narrow job it can do: reducing false positives without a large recall collapse. It is not a cure for the low raw recall. Large magenta gaps remain in the plots because Stage 5 never selected evidence there, and cleanup cannot create it. Stage 7 can connect supported observations with fitted geometry, but it should not be expected to recover lanes when the upstream anchors are absent. - **One decision worth looking at:** The direct component filter produces **17.831%** tolerant F1. The dash-chain restoration adds six components but no additional tolerant true-positive hit on this slice, leaving final F1 at **17.825%**. The chain behavior is implemented and tested, but the real-data ablation does not currently justify claiming it helps. Before checkpoint 10, I recommend deciding whether to validate the complete configured output or disable chain restoration and validate the slightly better direct-filter output. - **What I would check first:** Open the artifact index, then compare the best cleanup example with the worst one. In the plots, purple is what Stage 6 removed; the final error map tells you whether those removals were useful. The exact links and legend are in the next section. This checkpoint stops here. Checkpoint 10—the held-out validation and freeze decision—has not started. ## What you can inspect Start with the human-readable artifact index: - [full visual and numeric index](../artifacts/09_stage6_bev_cleanup/README.md); - [best observed F1 improvement](../artifacts/09_stage6_bev_cleanup/images/seq_04_001583943622370_stage6.png), from 23.06% to 25.13%; - [largest observed regression](../artifacts/09_stage6_bev_cleanup/images/seq_01_001487706805290_stage6.png), from 24.63% to 23.71%; - [complete per-frame table](../artifacts/09_stage6_bev_cleanup/benchmark.csv); - [machine-readable aggregate](../artifacts/09_stage6_bev_cleanup/benchmark.json); and - [per-frame component reports](../artifacts/09_stage6_bev_cleanup/reports/). The code entry points are [`bev_cleanup.py`](../src/lidar_lane_detection/bev_cleanup.py), [`bev_cleanup_diagnostics.py`](../src/lidar_lane_detection/bev_cleanup_diagnostics.py), and [`bev_cleanup_benchmark_cli.py`](../src/lidar_lane_detection/bev_cleanup_benchmark_cli.py). The synthetic behavior checks are in [`test_bev_cleanup.py`](../tests/test_bev_cleanup.py). ## How to read each six-panel plot Read panels 1 through 4 from left to right as an ablation: 1. **Raw Stage 5:** every green cell contains at least one raw global-Otsu candidate point. 2. **Longitudinal gap closing:** green survives unchanged; orange is a newly filled short gap in the same lateral column. 3. **Direct component filter:** green is retained; purple is rejected by the component shape, support, and strength policy. 4. **Final plus dash chains:** green passed directly; orange was restored only because compatible compact components formed an aligned forward chain. 5. **Ground truth:** magenta is a labeled K-Lane lane cell. The label is loaded only after prediction and never changes Stage 6. 6. **Strict error map:** white is an exact same-cell true positive, green is a false positive, and magenta is a false negative. For every BEV panel, the vehicle is below the image, forward is toward the top, and positive lateral `y` is to the left. The panel header reports tolerant F1 after every operation so you can tell which operation helped or hurt that specific frame. ## How to read the result tables - **Precision (P):** of what Stage 6 retained, how much is compatible with a labeled lane cell? - **Recall (R):** of the labeled lane cells, how many have nearby retained evidence? - **F1:** the balance between precision and recall. - **Strict:** prediction and label must occupy exactly the same cell. - **K-Lane tolerant:** a one-cell offset is accepted, matching K-Lane's confidence behavior; the outer border is not scored. - **TP / FP / FN:** true-positive, false-positive, and false-negative cell counts micro-summed across all successful frames. These are candidate-mask metrics, not final fitted-lane metrics. A candidate mask is sparse by design; Stage 7 is responsible for turning good anchors into continuous geometry. ## Exact operation-by-operation results ### K-Lane one-cell-tolerant metric | Output after operation | TP | FP | FN | Precision | Recall | F1 | |---|---:|---:|---:|---:|---:|---:| | Raw Stage 5 | 3,023 | 10,360 | 18,819 | 22.59% | 13.84% | 17.16% | | Short forward gap close | 3,023 | 10,442 | 18,819 | 22.45% | 13.84% | 17.12% | | Direct component filter | 2,977 | 8,572 | 18,865 | 25.78% | 13.63% | 17.83% | | Final plus dash chains | 2,977 | 8,583 | 18,865 | 25.75% | 13.63% | 17.83% | The gap close alone is deliberately small and slightly negative: it adds local continuity, not semantic evidence. Its value is that connected components can then measure a short broken structure as one object. The component filter is where the false-positive reduction occurs. The chain pass is neutral to slightly negative on this development sample. ### Strict same-cell metric | Output after operation | TP | FP | FN | Precision | Recall | F1 | |---|---:|---:|---:|---:|---:|---:| | Raw Stage 5 | 796 | 11,724 | 21,305 | 6.36% | 3.60% | 4.60% | | Short forward gap close | 797 | 11,807 | 21,304 | 6.32% | 3.61% | 4.59% | | Direct component filter | 779 | 9,872 | 21,322 | 7.31% | 3.52% | 4.76% | | Final plus dash chains | 779 | 9,883 | 21,322 | 7.31% | 3.52% | 4.76% | The large strict-versus-tolerant gap is expected for sparse LiDAR candidate cells and a 144-grid label. Both are reported so a tolerance improvement cannot hide exact-cell behavior. ## Per-sequence result | Sequence | Frames | Raw tolerant F1 | Final P | Final R | Final F1 | |---|---:|---:|---:|---:|---:| | `seq_01` | 10 | 24.71% | 34.33% | 19.81% | 25.12% | | `seq_02` | 10 | 15.42% | 39.74% | 9.65% | 15.53% | | `seq_03` | 10 | 18.42% | 28.35% | 14.50% | 19.19% | | `seq_04` | 9 | 8.10% | 8.02% | 10.12% | 8.95% | Stage 6 improves aggregate F1 in every sequence, but sequence 4 remains much weaker than the others. That agrees with checkpoint 08: sequence 4 often has poor lane-corridor support before cleanup begins. ## The implemented cleanup policy ### 1. High-resolution rasterization Raw candidate points are projected into `288 x 288` metric BEV cells. Each cell retains the maximum raw reflectivity and the number of candidate points. The mean frame contains 432.03 occupied high-resolution cells, which become 321.03 occupied cells on the `144 x 144` label grid. ### 2. Directed gap close Only gaps of at most two empty high-resolution cells are closed, and only when candidate evidence bounds the gap in the same lateral column. This bridges at most `0.32 m` in the forward direction. There is no lateral dilation and no six-cell-per-row budget. ### 3. Component measurements Eight-connected components are measured in metric coordinates. Every report contains: - supported cells and morphology-filled cells; - support density and candidate-point count; - forward span and lateral span; - PCA orientation and elongation; - centroid; - maximum reflectivity and its percentile among that frame's raw Stage-5 candidate points; and - whether the component passed directly, by a chain, or not at all. ### 4. Conservative component rejection The first geometry-only draft rejected locally transverse components. On the 39 readable development frames, it collapsed tolerant recall from 13.84% to 1.76% and F1 from 17.16% to 3.13%. Component analysis showed why: short real lane fragments often follow the LiDAR scan direction locally. Orientation is still measured and used to constrain chains, but it is no longer a blanket rejection rule. The selected development policy is: - keep singleton components because blanket singleton removal cost too many valid sparse anchors; - for components with two to four supported cells, reject only those whose maximum strength is in the bottom 10% of this frame's Stage-5 candidates; - for components with at least five supported cells and either more than `0.48 m` forward span or more than `0.64 m` lateral span, require maximum strength in the top 10% of the frame's Stage-5 candidates; and - preserve other compact components. That final top-decile test is not Stage 5.1. It applies only to a connected component already flagged as a long/wide clutter shape; it does not globally replace Otsu and has no per-row cell cap. ### 5. Dash-chain preservation A weak compact component may be restored only through a connected graph of at least three compact components whose centers are at most `3.0 m` apart, at most `0.24 m` laterally offset, and at most 15 degrees off the forward direction. The group must span at least `1.0 m` longitudinally. Synthetic tests cover aligned restoration, lateral misalignment, and direct-component anchoring. On this real development slice, only six of 6,766 total components are restored exclusively by this rule. They add 11 tolerant false-positive cells and no tolerant true-positive label hit. That is why the chain should receive special attention at the checkpoint-10 freeze decision. ## Development tuning and why it is not validation The configuration was selected after inspecting operation ablations on these 39 readable frames. The main candidates tested included strict geometry-only filtering, progressively wider fragment chains, compact-component strength gates, and long/wide-component strength gates. The strict geometry-only policy was rejected immediately because of its recall collapse. The final policy was chosen for conservative true-positive retention rather than maximum precision. Because configuration selection and these reported measurements use the same frames, the numbers are development evidence only. They are not an unbiased estimate of generalization. Checkpoint 10 exists to freeze the chosen policy first and then run a separate held-out sample. ## Runtime and output size Across 39 successful frames, Stage-6 cleanup alone takes: | Statistic | CPU time | |---|---:| | Median | 21.17 ms | | Mean | 27.68 ms | | 95th percentile | 59.81 ms | The final mask averages 364.18 occupied `288 x 288` cells and 273.38 occupied `144 x 144` cells per frame, versus 432.03 and 321.03 before cleanup. ## Failure handling `seq_04/001680832844300` declares 786,432 PCD values but contains only 160,567. The benchmark records this `KLaneFormatError`, continues with the remaining frames, and includes the failure in both the JSON and artifact index. This is the same malformed source frame seen in checkpoints 06 and 08. ## Verification The implementation passes all **34** repository tests. Six Stage-6-specific tests verify: - forward-only short-gap closing; - retention of a supported lane-like component; - rejection of a weak transverse clutter structure; - aligned compact-fragment chain restoration; - refusal to chain laterally misaligned fragments; - use of a direct component as a chain anchor; and - rejection of misaligned Stage-4/Stage-5 inputs. The source tree also passes Python bytecode compilation. Reproduce the development benchmark with: ```bash uv run klane-stage6 /mnt/arnavs_hdd/data/02-lane-detection/K-Lane \ --sequence 1,2,3,4 \ --sample-count 10 \ --output-dir artifacts/09_stage6_bev_cleanup ``` If the environment cannot reach PyPI but the existing virtual environment is already present, the equivalent offline command is: ```bash PYTHONPATH=src .venv/bin/python \ -m lidar_lane_detection.bev_cleanup_benchmark_cli \ /mnt/arnavs_hdd/data/02-lane-detection/K-Lane \ --sequence 1,2,3,4 \ --sample-count 10 \ --output-dir artifacts/09_stage6_bev_cleanup ``` ## Recommendation and stop gate Stage 6 is complete enough for review: code, tests, per-operation ablations, full component reports, development metrics, plots, legends, and failure cases are all present. My recommendation is to keep the conservative component filter, because it removes a meaningful number of false positives while retaining 98.48% of raw Stage-5 tolerant true positives. I would not yet freeze the dash-chain pass: its behavior is correct in controlled tests, but it provides no positive evidence on this development slice. Choose direct-filter output or full chain-restored output before checkpoint 10, freeze that choice, and then test it without retuning on held-out frames. Do not start Stage 7 until checkpoint 10 has made the Stage-6 freeze decision.