CHARD: Characteristic-Aware Hierarchical Detection

Trained checkpoints for "CHARD: Characteristic-Aware Hierarchical Detection for Bangladesh Road Scenes". Code: github.com/euler1729/chard.

CHARD turns the physical characteristics of a vehicle (wheel count, size, propulsion mechanism) into auxiliary supervision, letting rare, safety-critical vehicle classes — auto-rickshaws, cart vehicles, wheelchairs — borrow statistical strength from the more common classes that share their attributes. It's implemented in two forms: an end-to-end hierarchical attribute-then-class head on RT-DETR, and a lightweight auxiliary attribute branch on YOLO that's discarded before inference (zero deployment cost).

This repository holds the 79 checkpoints behind the paper's results: the primary single-run baselines plus the full 72-run multi-seed × multi-dataset study (8 model variants × 3 seeds × 3 Bangladeshi road-vehicle datasets — BadODD, Poribohon-BD, Sorokh-Poth).

Headline result (BadODD, test split, mean ± std over 3 seeds)

Model	mAP@[.5:.95]	Tail-class AP
YOLOv8-l	42.6%	28.7%
YOLOv10-l	39.4%	23.0%
YOLOv11-l	42.5%	28.0%
RT-DETR-l	41.1%	29.7%
CHARD-YOLOv11 (attr)	44.3%	31.9%

CHARD-YOLOv11 (attr) improves on its matched no-attribute control by +2.4 points (paired t-test, p = 0.003) at identical inference latency to the vanilla backbone. Full results: benchmarks/ in the code repo.

Files

badodd_seed0/           # primary single-run baselines + CHARD (seed 0), Table 5/6 of the paper
  yolov8l.pt  yolov10l.pt  yolo11l.pt  rtdetr_l.pt
  chard_rtdetr.pt  chard_yolov8l.pt  chard_yolov8l_attr.pt

badodd/                 # 72-run multi-seed study, Table 9 — 24 files per dataset dir:
poribohon_bd/           #   {yolov8l,yolov10l,yolo11l,rtdetr_l}_s{0,1,2}.pt
vehicle_data/           #   chard_{yolov8l,yolo11l}_{attr,noattr}_s{0,1,2}.pt
                         #   (vehicle_data = "Sorokh-Poth" in the paper)

All checkpoints are stripped of optimizer/scheduler/EMA state and stored in FP16 (see prepare_release_weights.py).

Loading

Plain YOLO baselines (yolov8l, yolov10l, yolo11l) load directly with Ultralytics:

from ultralytics import YOLO
model = YOLO("badodd_seed0/yolo11l.pt")
results = model.predict("image.jpg")

RT-DETR baseline (rtdetr_l) likewise:

from ultralytics import RTDETR
model = RTDETR("badodd_seed0/rtdetr_l.pt")

CHARD-YOLO checkpoints (chard_yolov8l*, chard_yolo11l*) are Ultralytics DetectionModel subclasses with an added attribute head — you need the chard_yolo package from the code repo registered before unpickling:

from ultralytics import YOLO
from chard_yolo.model import AttrDetectionModel  # noqa: F401 — required for unpickling
model = YOLO("badodd/chard_yolo11l_attr_s0.pt")

CHARD-RT-DETR (chard_rtdetr.pt) is a raw state_dict from a custom training loop (not an Ultralytics/Transformers checkpoint) — reconstruct the architecture from chard_model.py in the code repo, then:

import torch
ckpt = torch.load("badodd_seed0/chard_rtdetr.pt", map_location="cpu")
model.load_state_dict(ckpt["model"])

Citation

@article{hasan2026chard,
  title   = {{CHARD}: Characteristic-Aware Hierarchical Detection for {Bangladesh} Road Scenes},
  author  = {Hasan, Mahmudul and Fahad, Istiaq Ahmed and Arefin, Md Fahim and Khan, Md Mosaddek},
  journal = {IEEE Access},
  year    = {2026},
  note    = {In press}
}

License

MIT — see LICENSE in the code repo. Underlying dataset licenses (BadODD, Poribohon-BD, Sorokh-Poth) are governed by their original sources.