MonoPRIO: Monocular 3D Object Detection with Query-Adaptive Prior Conditioning

This repository hosts the official implementation of MonoPRIO, a model for monocular 3D detection with query-adaptive prior conditioning for robust size estimation under size-depth ambiguity. MonoPRIO builds upon MonoDETR and MonoDGP.

Our core idea is to stabilise ambiguous monocular size estimation with class-aware offline prior banks, query-level routed mixture priors, uncertainty-aware log-space size conditioning, and CAP regularisation in the size pathway.

KITTI Validation Results

The table below reports per seed validation results (R40) for our unified model and median of 5 values used in our paper. For baseline results of MonoDGP and MonoCLUE please refer here.

Run	Car APBEV|R40			Car AP3D|R40			Ped. AP3D|R40			Cyc. AP3D|R40			Log	Checkpoint
	E	M	H	E	M	H	E	M	H	E	M	H
Seed 444	37.7385	27.4488	24.7231	30.4130	21.9274	18.7024	12.4075	9.6523	7.4288	10.2545	5.3419	5.0021	log	ckpt
Seed 445	40.3231	28.7343	25.1377	31.0523	21.9635	19.5214	12.2332	9.3608	7.1981	14.9668	7.9881	7.0977	log	ckpt
Seed 446	38.3903	28.5380	24.7171	30.1636	21.7001	18.9655	10.9028	7.9890	6.3868	12.4020	5.9881	5.8987	log	ckpt
Seed 447	35.8443	26.6558	23.4343	27.4737	20.8672	18.0712	13.1752	9.9896	7.7053	10.9032	5.8934	5.1463	log	ckpt
Seed 448	40.6979	28.8151	25.0044	31.3979	21.8560	19.2577	12.7353	9.1760	7.3320	13.5628	6.7914	6.2966	log	ckpt
Median of 5 (paper)	38.390	28.538	24.723	30.413	21.856	18.965	12.408	9.361	7.332	12.402	5.988	5.899	-	-

KITTI Official Test Results

We also provide test set checkpoints and prediction files for direct verification on the KITTI 3D object detection benchmark.

Unified model

Model

Car APBEV|R40

Car AP3D|R40

Ped. AP3D|R40

Cyc. AP3D|R40

Checkpoint

Predictions

E

M

H

E

M

H

E

M

H

E

M

H

MonoPRIO

35.50

24.66

21.54

26.83

18.93

16.25

16.31

10.74

9.08

7.72

4.32

3.61

ckpt

pred

Car-only model

Model	Car APBEV|R40			Car AP3D|R40			Checkpoint	Predictions
	E	M	H	E	M	H
MonoPRIO	35.74	26.05	22.90	28.12	19.85	16.93	ckpt	pred

Test results submitted to the official KITTI Benchmark:

Car category:

All categories:

Baseline Validation Logs

For fairness and reproducibility, we also release the exact local validation logs used for the 5-seed median comparisons against MonoDGP and MonoCLUE.

MonoCLUE references:

• Paper
• GitHub

Method	Car APBEV|R40			Car AP3D|R40			Ped. AP3D|R40			Cyc. AP3D|R40
	Easy	Mod.	Hard	Easy	Mod.	Hard	Easy	Mod.	Hard	Easy	Mod.	Hard
MonoDGP Pu et al. (2025)	37.451	26.938	24.416	29.536	21.136	18.802	12.311	9.270	7.268	9.813	4.921	4.422
MonoCLUE Yang et al. (2025)	38.249	28.119	24.656	29.344	21.416	18.495	11.658	8.579	7.043	11.180	5.741	5.259
MonoPRIO	38.390	28.538	24.723	30.413	21.856	18.965	12.408	9.361	7.332	12.402	5.988	5.899

MonoDGP seed logs:

• seed444
• seed445
• seed446
• seed447
• seed448

MonoCLUE seed logs:

• seed444
• seed445
• seed446
• seed447
• seed448

Installation

1. Clone this project and create a conda environment:

   git clone https://github.com/bigggs/MonoPRIO.git
   cd MonoPRIO
   
   conda create -n monoprio python=3.10 -y
   conda activate monoprio

2. Install pytorch and torchvision matching your CUDA version:

   pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu130
   

3. Install requirements and compile the deformable attention:

   pip install -r requirements.txt
   
   cd lib/models/monoprio/ops/
   bash make.sh
   
   cd ../../../..

4. Download KITTI datasets and prepare the directory structure as:

   │MonoPRIO/
   ├──...
   │data/kitti/
   ├──ImageSets/
   ├──training/
   │   ├──image_2
   │   ├──label_2
   │   ├──calib
   ├──testing/
   │   ├──image_2
   │   ├──calib

5. Prepare prior banks

   Option A: download and use our priors

   Place downloaded files under MonoPRIO/priors/ then set model/prior_path in your config.

   Setup	Bank	Link
   Validation	Unified	Download
   Validation	Car-only	Download
   Test	Unified	Download
   Test	Car-only	Download

   Option B: generate priors locally from KITTI labels/images

   Install CLIP:

   pip install git+https://github.com/openai/CLIP.git

   Generate priors from the config split (defaults to dataset.train_split in configs/monoprio.yaml):

   python tools/build_priors.py --config configs/monoprio.yaml --out-dir priors

   To split a unified bank into per-class banks (ped/car/cyclist) (for if you want to just train on the car class):

   python tools/split_banks.py --input priors/priors_unified.npz --out-dir priors/individual_banks

   By default, training uses:

   model:
     prior_path: 'priors/priors_unified.npz'

   Update this path in your config if you want to use a different bank file.

Get Started

Train

You can modify the settings of models and training in configs/monoprio.yaml and indicate the GPU in train.sh:

bash train.sh configs/monoprio.yaml > logs/monoprio.log

Test

The best checkpoint will be evaluated as default. You can change it at "tester/checkpoint" in configs/monoprio.yaml:

bash test.sh configs/monoprio.yaml

You can test the inference time on your own device:

python tools/test_runtime.py

Citation

If you find our work useful in your research, please consider giving us a star and citing:

@misc{davies2026monoprioadaptivepriorconditioning,
      title={MonoPRIO: Adaptive Prior Conditioning for Unified Monocular 3D Object Detection}, 
      author={Leon Davies and Qinggang Meng and Mohamad Saada and Baihua Li and Simon Sølvsten},
      year={2026},
      eprint={2605.14781},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.14781}, 
}

Acknowlegment

This repo benefits from the excellent work MonoDETR, and MonoDGP. If you find this work useful, please also consider checking out and citing their work.