FAIR1M in oriented-det — 37 classes, and why 36.70% is not a failed train

FAIR1M asks the detector to name the object, not just find it: a Boeing or a C919, a small car or a van, 37 names in all.

A 12-epoch Faster R-CNN finetune reached 36.70% on the official val tiles. Published Faster R-CNN numbers on this set sit around 31–35%, so this run is in that band. The box often lands on the right object. The name is the hard part. How that shows up in the errors is below.

The dataset Link to heading

Sun et al., FAIR1M: A Benchmark Dataset for Fine-grained Object Recognition in High-Resolution Remote Sensing Imagery, ISPRS 2022. Official FAIR1M-1.0 has train 16,488 and val 8,287. Gaofen test labels are not public. Recipes train on train tiles only; val is not in training — unlike leaky DOTA eval-val.

Images are typically 1k–10k px. Do not train whole-image like HRSC. Convert, then tile.

odet fair1m-to-dota \
  --data-root /path/to/data/FAIR1M \
  --output-dir /path/to/data/FAIR1M-dota \
  --splits train,val

odet tile-dota /path/to/data/FAIR1M-dota/train --tile-size 1024 --overlap 200 --min-overlap 0.7
odet tile-dota /path/to/data/FAIR1M-dota/val   --tile-size 1024 --overlap 200 --min-overlap 0.7

odet train --config configs/rotated_faster_rcnn/fair1m_le90_1x.json

The 37-way classifier is randomly initialized (class-count tensors skipped on load). Notebook: notebooks/kaggle_fair1m_tutorial.ipynb — a 1-epoch smoke. It will not reproduce 36.7%.

Local 1× Faster R-CNN: 36.70% Link to heading

From hf://rotated_faster_rcnn_dota_le90_1x, NVIDIA L4, ~24.5 h (runs/rotated_faster_rcnn/20260910-072116). Score ≥ 0.05, rotated IoU 0.50, non-empty val tiles.

Not the Gaofen hidden test. Not FAIR1M mAP_F. No Hub slug.

The bottleneck is class ID, not boxes Link to heading

Epoch 12 mean best IoU vs any detection was 0.62, same-class 0.50, GT cover 62%. On DOTA, Faster R-CNN’s in-train GT-cover is about 88.6% (which detector to train). About 32.6k ground truths had IoU ≥ 0.5 with a wrong-class box. Train imbalance is 1038× (Small Car 143,249 vs C919 138). Rare subtypes stay near 0 AP. Sports fields are easy (Baseball Field 88.5%, Tennis Court 81% on this run).

Literature band (different test sets; cited as a band only): FAIR1M paper Faster R-CNN R101 31.53%; later Rotated Faster R-CNN R50 ~33–35%; Oriented R-CNN R50 ~39–42%.

The 1× recipe uses unweighted cross-entropy. Loss and mAP were still moving at epoch 12 (train loss 0.412, mAP50 36.70%), and the 37-way head started from random weights. These are the levers that target that, not a different dataset. None of them has a published FAIR1M number in this series.

Resume, or a 3× schedule, from this checkpoint. Twelve epochs is the 1× budget. A 3× run is the longer MMRotate-style schedule (36 epochs) starting from these weights, so the head is not re-initialized. The curve had not flattened. More epochs are the cheapest bet if the only problem is that training stopped early.

Oriented R-CNN 1× (configs/oriented_rcnn/fair1m_le90_1x.json). Same tiles, oriented proposals instead of a horizontal RPN. Literature for Oriented R-CNN R50 sits about 39–42%, against about 33–35% for Rotated Faster R-CNN. That gap is mostly better RoIs, not a smarter subtype head. It will not by itself name a C919, but it is why the upper end of the band is higher than 36.70%.

Coarse-to-fine groups (loss.roi_grouped_ce_*). FAIR1M’s five groups are ship, vehicle, airplane, court, and road (FAIR1M_GROUPS). Early epochs can train those coarse labels while the head stays 37-way: a Boeing737 called a C919 is not an error while both are “airplane.” A step or linear_ramp schedule then hands the loss back to the fine names. This needs loss_type cross_entropy or class_weighted.

Class weights, or focal loss, for the 1038× imbalance. Inverse-frequency weights (class_weighted, or class_weight_*) stop Small Car (143,249) from owning the gradient over C919 (138). Focal loss (loss_type: focal) does the same job a different way: easy, frequent boxes contribute less, so rare subtypes are not washed out. focal_weighted applies the class weights on top of focal. Focal does not combine with grouped CE — the trainer ignores roi_grouped_ce_* when the loss is focal. Pick a curriculum (groups, then fine labels) or a reweighting (class weights and/or focal), not both at once.

Links Link to heading

• oriented-det docs — FAIR1M · v0.3 release note
• FAIR1M paper (arXiv)
• Previous: Oriented-Det v0.3.1 · HRSC2016
• Next: SSDD (28 Sep)

Written on September 24, 2026 by Jeff Faudi. Link to heading