[犀牛鸟-A2]：Add configurable scale-aware TAL assignment for small objects - #274
[犀牛鸟-A2]：Add configurable scale-aware TAL assignment for small objects#274YilinWang121054 wants to merge 1 commit into
Conversation
Consolidated onto upstream 7bfbfd3. The complete tree matches local tested merge 6f55860. Training evidence retains its original source commits.
|
CI follow-up (2026-09-11): I compared the failing Windows and Ubuntu Python 3.13 jobs against the exact upstream base, 7bfbfd3. Both shards have the same failing test names and error summaries as the base.
The shared failures are the foundation configuration error-message assertion and cached-response distillation's feature/stride count mismatch. Windows also has the offline-cache key validation failure. Although one traceback points to Comparison, original job links, preserved logs and local reproduction. The local Windows sandbox blocks the offline test's attempted root-directory write, so its final exception differs from hosted CI; this is documented separately. I have not skipped these tests or changed foundation behavior to make the checks green. This note explains the two checked failures; it does not claim that all CI passes. |
|
P0 评测材料补充（2026-09-11） 补交一份官方 DET 源码与 Python 实现的逐项对照，参考版本锁定为 VisDrone2018-DET-toolkit 的 使用原始 P0 第120轮 报告保留了 adaptive 的负结果。seed 20260824 因 auto/momentum 差异单列为探索性，早期缺完整逐轮统计的组仅供参考，不用于正式三seed的 P1 达标判定。原始基线的完整配置、checkpoint SHA-256、120轮统计和原始日志都已公开。如果源码对应关系还有遗漏，欢迎指出，我会继续补充。 |
2 participants
Summary
Small VisDrone objects can contain too few anchor centers for TAL, even with a larger top-k. This change adds an explicit
stal_mode=tal|fixed|adaptiveswitch. Adaptive mode expands the candidate region of small GT boxes, adds nearby candidates when coverage is insufficient, and uses separate top-k values for small, medium, and large objects.The configuration lives in
default.yaml, with checks for types, finite values, ranges, and parameter relationships. The default remains fixed-stride expansion. Alpha, beta, and the IoU ranking metric are unchanged. The candidate-conflict fix is already merged in #253 and is a dependency of this work, not a second bug-fix contribution.Tests
The review checkout, merged with upstream main at 7bfbfd3, passes the STAL, default-configuration, model-configuration and TAL conflict tests on CPU:
Coverage includes candidate geometry, area boundaries, empty GT, tiny boxes, overlapping targets, candidate conflicts, configuration injection and invalid parameters (including NaN/Inf). MPS is unavailable in this run. The autocast test exercises a small assignment/loss fixture; this CPU run does not certify real-batch CUDA AMP parity. The changed files retain 15 Ruff diagnostics already present in current main, with no new diagnostics; their formatting and spelling checks pass. Repository-wide quality checks are not clean, and their outputs are retained separately.
Experiments
The full-data experiments use YOLO-Master v0.1-N, VisDrone2019-DET, imgsz=800, batch=4, FP32, Mosaic on, and 120 epochs. The primary checkpoint is
last.ptfrom epoch 120. APs uses original-image GT area <32² and maxDets=500, after the VisDrone ignore-region filter. These area bins are an A2 supplement.These results do not meet the A2 +1.0 APs target. Seed 20260824 also differs from later seeds in actual optimizer momentum (0.937 versus 0.9), so the rows are not presented as a uniform three-seed aggregate. A fixed-subset smoke reduced small-GT zero-positive rates, but this has not translated into higher APs in the completed pairs.
The locked acce839 P0 baseline has also completed 120 epochs and full 548-image validation. Its supplemental APs/APm/APl are 12.5117/30.7556/39.7667. All 120 online assignment records were checked against the original files and cover 194,160 training batches. This baseline is recorded separately from the fixed comparison, which includes #253.
Research report (Chinese) · P0 metrics, predictions and 120 assignment records · Evidence index · Regression test log. These links are pinned to the verified evidence commit. This PR does not claim an accuracy improvement or P1 acceptance.
Limitations
The minimum candidate count is enforced before conflict resolution; it does not guarantee that every GT retains that many positives. Candidate counts also do not measure normalized target weights. Full online per-epoch statistics are available for the new P0 baseline, not retrospectively for the older TAL/fixed/adaptive runs. The remaining paired runs and Mosaic interaction are incomplete.
The original MATLAB devkit could not run on this host because MATLAB R2024a reports MathWorks error 5201. The review package includes a MATLAB-source audit port and concrete prediction-level counterexamples for the older Python port; its runtime parity remains pending until the original tool runs on a licensed host. P2 threshold/warmup scans and a second dataset/task have not been completed. The latest-main diff was checked to exclude the already merged #253 fix.