badas 1.1.2

Beyond ADAS — collision anticipation inference for dashcam video

BADAS - Beyond ADAS

BADAS (Beyond ADAS) is a deep learning framework for predicting collision likelihood in dashcam video sequences. It supports multiple vision transformer backbones and deployment formats optimized for real-time inference.

Installation

pip install badas

Optional extras for faster video decoding or alternative backends:

pip install decord           # Fast video decoding (falls back to OpenCV)
pip install onnxruntime-gpu  # ONNX backend

Quick Start

from badas.inference import BADAS, BADASConfig

predictor = BADAS.from_pretrained("nexar-ai/BADAS-Open")

# Batch prediction on a video
results = predictor.predict_video("dashcam.mp4", stride=1)
for r in results:
    print(f"[{r['timestamp']:.2f}s]  {r['risk_level']:<6}  p={r['probability']:.3f}")

# Streaming / moving-window inference
for pred in predictor.predict_stream("dashcam.mp4", stride=1):
    if pred['probability'] > 0.7:
        print(f"WARNING: High collision risk at {pred['timestamp']:.2f}s")

Prediction Output

Each prediction is a dict:

{
    'frame_index': 100,
    'timestamp': 4.17,
    'probability': 0.823,
    'risk_level': 'high',   # 'low' | 'medium' | 'high'
    'smoothed': True,
}

Inference Backends

Format	Load	Notes
.ckpt / .pt	BADAS("model.ckpt")	Full model + config, supports torch.compile
.onnx	BADAS("model.onnx")	CPU/GPU via ONNX Runtime
.trt / .engine	BADAS("model.trt")	NVIDIA GPU, lowest latency

Configuration

from badas.inference import BADAS, BADASConfig, SmoothingConfig

config = BADASConfig(
    use_compile=True,       # torch.compile — slow first run, fast thereafter
    startup_ramp=True,      # yield predictions before the full window is buffered
    smoothing=SmoothingConfig(
        enabled=True,
        alpha_rise=0.7,     # smoothing when risk increases
        alpha_fall=0.3,     # smoothing when risk decreases
    ),
)
predictor = BADAS("model.ckpt", config=config)

Available Models

Open Source

Model	Size	Notes
nexar-ai/BADAS-Open	ViT-L	Publicly available

predictor = BADAS.from_pretrained("nexar-ai/BADAS-Open")

Commercial (Nexar)

Model	Params	Architecture	Notes
nexar-ai/badas-2.0	330M	ViT-L	Best accuracy
nexar-ai/badas-2.0-flash	86M	ViT-B	Fast, default
nexar-ai/badas-2.0-flash-lite	22M	ViT-S	Fastest
nexar-ai/badas-1.0	330M	ViT-L	Previous version

To request access, visit nexar-ai.com/badas.

Inference API

For production use, BADAS is available as a hosted REST API at https://badas-api.corp.nexars.ai. The BADASClient handles uploads, async job queuing, SSE streaming, and progress bars.

Installation

requests and tqdm are included in the base badas package — no extra install needed.

Setup

from badas.client import BADASClient

client = BADASClient(api_key="nxr_live_...")

# Check that models are ready
health = client.health()
print(health["status"], health["models_loaded"])

Predict (blocking)

Submits an async job and polls until complete. Returns a list of prediction dicts.

preds = client.predict("dashcam.mp4")
for p in preds:
    print(f"[{p['timestamp']:.2f}s]  {p['risk_level']:<6}  p={p['probability']:.3f}")

Stream (window-by-window)

Yields each prediction as soon as its window is processed. Reconnects automatically if the gateway drops the connection.

for pred in client.stream("dashcam.mp4"):
    print(f"[{pred['timestamp']:.1f}s] {pred['risk_level']} ({pred['probability']:.0%})")

Async job (submit / poll / stream)

# Submit returns a job ID immediately — processing runs in the background
job_id = client.submit("dashcam.mp4")

# Stream results as they arrive
for pred in client.stream_job(job_id):
    ...

# Or wait for completion and get all results at once
preds = client.poll_job(job_id)

Inference options

Parameter	Default	Description
model	badas-2.0-flash	Model ID (see table above)
stride	1	Window stride in model-fps frames
threshold	0.75	Risk score that triggers heatmaps / VLM descriptions
heatmaps	"none"	"none" | "threshold" | "always"
describe	False	VLM hazard description on trigger windows (auto-sets heatmaps="threshold")
include_maps	False	Include raw hazard_map / attention_map arrays
progress	True	Show tqdm progress bar (adapts to notebook / CLI)

All options are available on predict, stream, and submit.

Heatmaps and VLM descriptions

import numpy as np

for pred in client.stream(
    "dashcam.mp4",
    describe=True,    # VLM hazard description for high-risk windows
    threshold=0.8,    # trigger threshold for heatmaps + descriptions
    include_maps=True,
):
    if pred.get("hazard_map") is not None:
        pred["hazard_map"] = np.array(pred["hazard_map"])  # H×W float array
    if pred.get("description"):
        print(f"[{pred['timestamp']:.1f}s] {pred['description']}")

Large files (> 50 MB)

Files above 50 MB are automatically uploaded to GCS via a signed URL before inference — no changes needed in calling code.

Model comparison

for model_id in ["nexar-ai/badas-2.0", "nexar-ai/badas-2.0-flash", "nexar-ai/badas-2.0-flash-lite"]:
    preds = client.predict("dashcam.mp4", model=model_id)
    peak  = max(p["probability"] for p in preds)
    print(f"{model_id}: {len(preds)} windows, peak={peak:.3f}")

Available models

Model	Params	Architecture	Notes
nexar-ai/badas-2.0	330M	ViT-L	Best accuracy
nexar-ai/badas-2.0-flash	86M	ViT-B	Fast, default
nexar-ai/badas-2.0-flash-lite	22M	ViT-S	Fastest
nexar-ai/badas-1.0	330M	ViT-L	Previous version

Citation

@article{goldshmidt2025badas,
  title={BADAS: Context Aware Collision Prediction Using Real-World Dashcam Data},
  author={Goldshmidt, Roni and Scott, Hamish and Niccolini, Lorenzo and
          Zhu, Shizhan and Moura, Daniel and Zvitia, Orly},
  journal={arXiv preprint},
  year={2025}
}