FlyLoom-v1: Event-Camera Looming Detection Benchmark

The first standardized benchmark for biologically-inspired looming detection.

Dataset Description

FlyLoom-v1 simulates a fly's eye watching approaching objects. The task: detect looming threats (predators on collision course) and trigger an escape reflex — before impact.

This benchmark was developed alongside FlyReflexNet, a connectome-constrained spiking neural network that matches standard ML accuracy at 124× fewer operations.

Key Features

Property	Value
Total samples	11,000
Input shape	(T=10, C=2, H=32, W=32)
Input type	Binary event tensor (ON/OFF polarities)
Event sparsity	~3% per frame (matches real DVS cameras)
Task	Binary classification + timing
Classes	5 stimulus types, 2 labels

Splits

Split	Samples	Threats	Non-threats
train	8,000	3,236	4,764
val	1,000	410	590
test	2,000	760	1,240

Stimulus Classes

Based on the classical Drosophila escape paradigm (Card & Dickinson 2008):

Class	Label	Description
Threat	1	Dark disc looming on collision course → escape required
Distractor	0	Dark disc looming but veering away → tests specificity
Translating	0	Dark disc moving sideways at constant size
Receding	0	Dark disc shrinking (moving away)
Static	0	Dark disc not moving

Usage

With Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("mayank171717/FlyLoom-v1")

# Access splits
train = ds["train"]
val = ds["validation"]
test = ds["test"]

# Get a sample
sample = train[0]
events = sample["events"]   # (10, 2, 32, 32) uint8
label = sample["label"]     # 0 or 1
impact = sample["impact"]   # step when predator strikes (-1 if non-threat)
kind = sample["kind"]       # "threat", "distractor", etc.

With PyTorch (Direct)

import numpy as np
import torch
from torch.utils.data import Dataset

class FlyLoomV1(Dataset):
    def __init__(self, split="train"):
        base = "path/to/FlyLoom_v1"
        self.events = np.load(f"{base}/{split}/events.npy")
        self.labels = np.load(f"{base}/{split}/labels.npy")
        
    def __len__(self):
        return len(self.labels)
    
    def __getitem__(self, idx):
        return (
            torch.from_numpy(self.events[idx].astype(np.float32)),
            int(self.labels[idx])
        )

The Task

Given a 10-step event stream, output a binary fire decision:

• 1 = JUMP (looming threat detected)
• 0 = HOLD FIRE (safe)

Timing matters. For threats, the model should fire before impact_step. Later detection = lower survival.

Metrics

Metric	Description	Goal
Survival rate	% of threats detected before impact	↑
False alarm rate	% of non-threats that triggered escape	↓
Reaction time	Mean steps-to-fire × 10 ms	↓
Operations/sample	Computational cost	↓

Benchmark Results

Model	Survival ↑	False Alarm ↓	Ops/sample ↓
Random	50%	50%	0
Flat-pixel GRU	96.7%	0.8%	4,059,530
GRU + retinal encoder	98.8%	0.1%	408,310
Spiking Fly (Phase 7)	96.2%	13.8%	1,458
Spiking Fly + Neuromod (ours)	99.3%	0.6%	3,296

Our model matches GRU accuracy with 124× fewer operations.

Biological Grounding

This dataset reproduces stimuli from real Drosophila escape experiments:

• Gabbiani et al. (1999) — The η = l/|v| computation in LGMD neurons
• Card & Dickinson (2008) — Visually mediated motor planning in fly escape
• von Reyn et al. (2014) — Giant Fiber pathway for escape triggering

Our model's reaction times show the same qualitative dependence on the r/v looming parameter as real flies.

Dataset Generation

Fully reproducible with fixed seed:

# The exact parameters used (from manifest.json)
{
    "seed": 42,
    "T": 10,
    "H": 32,
    "W": 32,
    "event_threshold": 0.08,
    "shot_noise": 0.001,
    "class_weights": {
        "threat": 0.4,
        "translating": 0.15,
        "receding": 0.15,
        "static": 0.15,
        "distractor": 0.15
    }
}

File Structure

FlyLoom_v1/
├── manifest.json          # Full config + stats
├── train/
│   ├── events.npy         # (8000, 10, 2, 32, 32) uint8
│   ├── labels.npy         # (8000,) int8
│   ├── impact.npy         # (8000,) int8
│   └── kinds.json         # list of 8000 strings
├── val/
│   └── ...
└── test/
    └── ...

Citation

@dataset{flyloom2025,
  author = {Mewar, Mayank},
  title = {FlyLoom-v1: Event-Camera Looming Detection Benchmark},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/mayank171717/FlyLoom-v1}
}

If you use this for biological comparisons, also cite:

@article{card2008,
  title={Visually mediated motor planning in the escape response of Drosophila},
  author={Card, Gwyneth and Dickinson, Michael H},
  journal={Current Biology},
  year={2008}
}

@article{gabbiani1999,
  title={Computation of object approach by a wide-field motion-sensitive neuron},
  author={Gabbiani, Fabrizio and Krapp, Holger G and Laurent, Gilles},
  journal={Journal of Neuroscience},
  year={1999}
}

License

CC-BY 4.0 — Free to use, remix, and redistribute with attribution.

Links

• Code: GitHub repository (update with your actual repo)
• Paper: (coming soon)

Contact

For questions or collaborations, open an issue on GitHub or contact via Hugging Face.