[ErshovVE]

Motivation

RecMetric reports one accuracy and one normalized edit distance over all lines. A model that reads short lines
well and long ones badly looks fine in that number. This is how a training problem on long text lines (CTC length,
see #18420) stayed hidden in our runs. The training log and tools/eval.py give no way to see accuracy by line
length.

Change

New optional RecMetric argument length_buckets: increasing upper bounds of the label length.

Metric:
  name: RecMetric
  main_indicator: acc
  length_buckets: [15, 40, 60]   # buckets 0-15, 16-40, 41-60, 61+

For every non-empty bucket the metric dict gets three more keys:

• acc_len_<lo>_<hi>: line accuracy;
• ned_len_<lo>_<hi>: 1 − normalized edit distance;
• n_len_<lo>_<hi>: number of lines;

for example acc_len_0_15, acc_len_61_inf. They appear in the training log and in tools/eval.py like the
existing keys.

• Without length_buckets the metric dict is exactly as before; main_indicator and best-model selection do not
  change.
• Upper bounds are inclusive. The length is that of the label as compared, so with ignore_space: True (the default)
  spaces are not counted, and with is_filter: True the filtered label is used.
• Empty buckets are not reported.
• A non-increasing or empty list fails at construction.
• length_buckets is passed through **kwargs like the other RecMetric options, so the metric wrappers that build
  RecMetric from config get it too.

Example from tools/eval.py (PP-OCRv6 small, 4 241 real Russian/English document lines, ignore_space: False):

acc: 0.9599, norm_edit_dis: 0.9895,
acc_len_0_15: 0.9358, ned_len_0_15: 0.9668, n_len_0_15: 1246.0,
acc_len_16_40: 0.9697, ned_len_16_40: 0.9981, n_len_16_40: 1024.0,
acc_len_41_60: 0.9737, ned_len_41_60: 0.9994, n_len_41_60: 1257.0,
acc_len_61_inf: 0.9636, ned_len_61_inf: 0.9994, n_len_61_inf: 714.0

Tests

tests/ppocr/test_rec_metric_buckets.py (15 tests):

• bucket names and the bucket of a given length, bounds inclusive;
• invalid lists rejected;
• per-bucket accuracy, edit distance and counts against hand-computed values;
• empty buckets left out;
• the dict without length_buckets unchanged;
• the length counted after ignore_space;
• reset() between evaluations.