MonoPCA

Rethinking Geometric Depth in Monocular 3D Object Detection: A Projection-Consistent Reformulation (NeurIPS 2026)

Zhihao Zhang, Abhinav Kumar, Huaizhi Qu, Tianlong Chen, Xiaoming Liu

Code | Project Page

This repository hosts the pretrained MonoPCA model with the MoB (Mixture of Backbone) backbone. MoB fuses a ResNet-50 pyramid with EVA-02-L, and depth is predicted with PCDepth. This model serves as the teacher for GeoAlign.

Model

Backbone	MoB: ResNet-50 + EVA-02-L
Training data	KITTI + nuScenes (front camera) + Waymo (front camera)
Classes	Car, Pedestrian, Trailer, Bus, Motorcycle, Cyclist, Barrier, Truck, Van, Traffic_cone, Bicycle
Input	1280 x 512. Each image is resampled to a virtual focal length of 748.84 px.
Parameters	354M

Files

• model.safetensors: model weights.
• monopca_inference.yaml: inference-only config. It is the same file as configs/monopca_inference.yaml in the code repository.

Usage

Set up the code repository, then compile the deformable attention op:

git clone https://github.com/alanzhangcs/MonoPCA.git
cd MonoPCA
pip install -r requirements.txt
cd lib/models/monopca/ops && bash make.sh && cd ../../../..

Run inference. The weights are downloaded from this repository automatically:

# a single image with its KITTI calib file (P2 is used)
python tools/inference.py --image path/to/000001.png --calib path/to/000001.txt --output outputs/inference

# a folder of images with calib files of the same name
python tools/inference.py --image path/to/image_2 --calib path/to/calib --output outputs/inference

Each output file <image_name>.txt holds one object per line in KITTI format, with a confidence score at the end:

type truncated occluded alpha x1 y1 x2 y2 h w l x y z ry score

The 3D boxes are in the camera coordinate system of the given calibration.

You can also load the model in Python:

import yaml
from PIL import Image
from tools.inference import load_model, predict
from lib.datasets.kitti.kitti_utils import Calibration

cfg = yaml.safe_load(open("configs/monopca_inference.yaml"))
model = load_model(cfg)  # downloads model.safetensors from zhihao406/MonoPCA
preds = predict(model, Image.open("000001.png").convert("RGB"), Calibration("000001.txt"), cfg)

A CUDA GPU is required, because the multi-scale deformable attention op only runs on CUDA.

Citation

@inproceedings{zhang2026rethinking,
  title     = {Rethinking Geometric Depth in Monocular 3D Object Detection:
               A Projection-Consistent Reformulation},
  author    = {Zhang, Zhihao and Kumar, Abhinav and Qu, Huaizhi and
               Chen, Tianlong and Liu, Xiaoming},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}