CDC CONTEST 2026

Contest

Crime Detection Challenge (CDC26) is an international competition organized to encourage participants to develop advanced methods for detecting crime events in real-world video. The challenge considers a wide range of situations found in surveillance applications, ranging from public spaces to semi-private environments such as retail stores, care facilities, and correctional institutions, where scene variability, occlusions, viewpoint changes, and heterogeneous video quality make reliable detection difficult.

Crime detection is a key requirement for modern security and a central problem in video analytics for surveillance. This capability is increasingly relevant for smart-city deployments and infrastructure monitoring. It is challenging because systems must recognize a wide variety of events (e.g. fighting, shooting, stealing) in videos that are mostly normal, while truly anomalous behaviors are rare, diverse, and often visually similar to benign actions. As a result, current models can be fragile: their predictions may become unstable, increasing false alarms and effort for human operators and their accuracy may drop when conditions differ from training data, leading to poor generalization under domain shifts. Progress therefore depends on models that separate normal from abnormal behavior for the right reasons, avoiding spurious correlations and overfitting to dataset biases that can affect performance across contexts. CDC26 makes these difficulties measurable and comparable through a shared benchmark that rewards robustness and generalization, a necessary step toward methods that work reliably in real scenarios.

Task definition and expected outcomes

Task: video-level crime detection with detection instant

Given an input video clip, participants must determine whether the clip contains a crime event (positive) or represents normal activity (negative).

For each video, the method must output:

a binary decision: crime / normal;
a detection instant p: frame index at which the method triggers the crime alarm, intended to correspond to the onset of the crime event.

In the test set, for negative videos, no crime event is present; therefore, any alarm raised by the method is counted as a false positive. For positive videos, each clip contains one crime event only, annotated with a start time and end time ($g_{start}$, $g_{end}$). The method must output a detection instant p, which is intended to correspond to the onset of the event and will be evaluated with respect to the annotated temporal range under the challenge protocol.

MIVIA-CDC Dataset

CDC26 is based on a balanced collection of real-world surveillance video clips derived from UCF-Crime, extended with precise temporal annotations marking when the anomalous event begins (and its temporal extent). The dataset covers 13 anomalous activities plus Normal events. Anomalous activities include: Abuse, Arrest, Arson, Assault, Burglary, Explosion, Fighting, Road Accident, Robbery, Shooting, Shoplifting, Stealing, Vandalism.

1.900

Total Videos

950

Normal

950

Anomalous

Temporal Annotations

Provided temporal annotations to support training and calibration.

Test set

The test set is completely new and private, containing 270 videos (135 anomalous + 135 normal) and will not be released to participants.

Challenging Negative Samples

The normal videos of the test-set include hard negative samples: realistic scenes where one or more anomalies could plausibly occur (e.g., roads for accidents, offices for arrests, parking lots for explosions/shootings, shops for stealing/shoplifting), but no crime actually happens, to stress false-alarm robustness.

Evaluation Protocol and Metrics

The crime detection accuracy of the competing methods will be evaluated in terms of Precision, Recall, and F1-Score. To formalize these metrics, it is necessary to define the sets of true positives $TP$, false positives $FP$, and false negatives $FN$.

Our test set contains both positive and negative samples (i.e., video depicting crime events or not, respectively). Each positive video shows one crime event with a start time and end time ($g_{start}$, $g_{end}$). Each prediction is evaluated by comparing the anomaly start instant $g_{start}$ with the detection one $p$. We indicate as the anomaly start instant the first frame in which it is visible. Each method must output a detection instant $p$ (frame index) for each test video, corresponding to the moment at which the method triggers the crime alarm.

Consequently, we define:

True Positives $TP$ all the detections occurring in positive videos at $g_{start} \le p \le g_{end}$.
False Positives $FP$ all the detections occurring at any index $t$ in negative or positive videos at $p < g_{start}$ or $p > g_{end}$.
False Negatives $FN$ the set of positive videos for which no valid detection occurs.

Defined these sets for $TP$, $FP$ and $FN$, we can compute the Precision ($P$), Recall ($R$) and F-Score ($F$) with respect to the number of true positives $|TP|$, false positives $|FP|$ and false negatives $|FN|$:

Precision

Measures the system's ability to reject false alarms. Higher precision indicates fewer false positives.

$P = \frac{| T P |}{| T P | + | F P |}$

$P$ assumes values in the range $[0,1]$ and measures the capability of the methods to reject $FP$; the higher is $P$, the higher the reliability of the method when it detects a positive sample.

Recall

Measures the system's sensitivity to detecting crimes. Higher recall means fewer missed events.

$R = \frac{| T P |}{| T P | + | F N |}$

$R$ assumes values in the range $[0,1]$ and evaluates the sensitivity of the method to detect crime; the higher is $R$, the higher the sensitivity of the method.

F1-Score

Provides a balanced measure combining precision and recall. The harmonic mean of both metrics.

$F_{1} = 2 \times \frac{P \times R}{P + R}$

$F$ assumes values from $[0,1]$ and measures the balance between precision and recall.

Final Ranking: The contest winner will be determined based on the highest F1-Score achieved on the test set, as it provides the most balanced evaluation of detection accuracy combining both precision and recall performance.

However, to emphasize the applicability of these methods in real-time scenarios, we also compute additional metrics. To find the model with the best performance and the most efficient solution for real-time systems, we consider measures that evaluate both detection speed and computational efficiency. These metrics help identify methods that can operate effectively in resource-constrained surveillance environments while maintaining high detection accuracy.

Notification Delay

Measures how quickly the method detects crimes after it occurs. Lower notification delay means faster response, crucial for timely intervention in crimes. The normalized score transforms the average delay into a value between 0 (worst performance, maximum delay) and 1 (best performance, instantaneous detection).

Variables:

$ d_i = |p_i - g_{start-i}| $ (detection delay for video)

$p_{i}$ = detection instant
$ g_{start-i} $ (ground truth instant)

$D = \frac{\sum_{i = 1}^{| T P |} d_{i}}{| T P |}$ (average delay in seconds)
$ T_{max}$ = 30 seconds

$D_{n o r m} = max (0, 1 - \frac{D}{T_{m a x}})$

Processing Frame Rate

The average number of frames processed by the method in one second on a target GPU. Higher PFR indicates faster processing speed, essential for real-time surveillance applications.

Variables:

$P F R_{t a r g e t}$ = target processing frame rate;
$P F R = \frac{N}{\sum_{i = 1}^{N} t_{i}}$ (frames per second);

$t_{i}$ = processing time for frame $i$ (seconds);
$N$ = total number of frames processed;

$P F R_{d e l t a} = max (0, \frac{P F R_{t a r g e t}}{P F R} - 1)$

Memory Usage

We measure the memory usage MEM, namely the memory in GB occupied by the method on the target GPU. The lower is MEM, lower is the necessary memory on the processing device

Variables:

$M E M_{t a r g e t}$ = target memory threshold;
$M E M$ = peak GPU memory consumption (GB);

$M E M_{d e l t a} = max (0, \frac{M E M}{M E M_{t a r g e t}} - 1)$

Rules

The deadline for the submission of the methods is 19th May, 2026. The submission must be done with an email at mdb2026@unisa.it in which the participants share (directly or with external links) the trained model, the code and the report. Please follow the detailed instructions reported here.
The participants can receive the training set and its annotations by filling out this form.
The participants can use these training samples and annotations for training and validating their methods, but they are also encouraged to use other external data for improving the performance of their models.
The participants must submit their trained model and their code by carefully following the detailed instructions reported here.
The participants are strongly encouraged to submit a contest paper to MDB 2026, whose deadline is 12th June, 2026. Authors can find complete instructions of how to format their papers in the workshop website.

WORKSHOP WEBSITE

Instructions

The methods proposed by the participants will be executed on a private test set. To leave the participants totally free to use all the software libraries they prefer and to correctly reproduce their processing pipeline, the evaluation will be done on Google Colab (follow this tutorial) by running the code submitted by the participants on the samples of our private test set.

Therefore, the participants must submit an archive including the following elements:

A Python script test.py, which takes as input the folder of the test videos (--videos) and produces as output a .CSV file, contained in the folder --results, which can be specified by the user.
The .CSV file must have one column containing the video name and one column containing crime start instant (frame index). If no crime is detected in the video the start column must be left empty.
Thus, the script may be executed with the following command:
python test.py --videos foo_videos/ --results foo_results/
A Google Colab Notebook test.ipynb, which includes the commands for installing all the software requirements and executes the script test.py.
All the files necessary for running the test, namely the trained model, additional scripts and so on.
The provided sample test.py includes the writing of the file with the results.

The submission must be done by email at mdb2026@unisa.it. The archive file can be attached to the e-mail or shared with external links. We strongly recommend to follow the example of code to prepare the submission.

Results

The final results of the CDC Challenge will be published here after the evaluation of all submitted methods on the private test set. Stay tuned for updates on the leaderboard and detailed performance metrics of each method!

ORGANIZING COMMITTEE

VINCENZO CARLETTI

Department of Information Engineering, Electrical Engineering, and Applied Mathematics (DIEM)
University of Salerno, Italy

vcarletti@unisa.it

ANTONIO GRECO

Department of Information Engineering, Electrical Engineering, and Applied Mathematics (DIEM)
University of Salerno, Italy

agreco@unisa.it

MATTIA MARSEGLIA

Department of Information Engineering, Electrical Engineering, and Applied Mathematics (DIEM)
University of Salerno, Italy

mmarseglia@unisa.it

hosted by

Crime Detection Challenge

August 31th, 2026

EVENT STARTS IN

CDC CONTEST 2026

Contest

Task definition and expected outcomes

Task: video-level crime detection with detection instant

MIVIA-CDC Dataset

Total Videos

Normal

Anomalous

Temporal Annotations

Test set

Challenging Negative Samples

Evaluation Protocol and Metrics

Precision

Recall

F1-Score

Notification Delay

Processing Frame Rate

Memory Usage

Rules

Instructions

Results

ORGANIZING COMMITTEE

VINCENZO CARLETTI

ANTONIO GRECO

MATTIA MARSEGLIA