Cross domain additive manufacturing video object segmentation dataset

Abstract

This article presents labeled data collected from video footage of additive manufacturing processes. This computer vision dataset contains 90 labeled additive manufacturing video segments across four distinct additive manufacturing processes, amounting to 900 labeled, object instance video frames. The technologies covered by the dataset include laser hot-wire additive manufacturing (LHW-DED), plasma arc welding (PAW), tungsten inert gas wire arc additive manufacturing (TIG-WAAM), and polymer-based additive manufacturing (visPolymer and irPolymer). For each of the manufacturing processes, build deposition video was captured and then randomly sampled into 10 frame-long video segments. Depending on the manufacturing process, two of the following four object instance classes are labeled in each frame: Melt Pool, Feed Wire, Nozzle, and Material. Since the dataset’s directory organization follows that of common video object segmentation (VOS) datasets, such as coMplex video Object SEgmentation (MOSE), MOSEv2, Densely Annotated Video Segmentation (DAVIS), and YouTube-VOS, it can be easily integrated into other VOS model training pipelines for foundation model fine-tuning, training, or inference capability testing. This dataset provides researchers access to labeled additive manufacturing object instances for VOS tasks.

Specifications Table

Subject	Computer Sciences
Specific subject area	Additive Manufacturing Video Object Segmentation
Type of data	Raw and Labeled Video Frames (.jpg and .png)
Data collection	Data was collected by researchers at Oak Ridge National Laboratory’s Manufacturing Demonstration Facility. The LHW-DED build was deposited using a Mazak VTC-800 G SR AM hybrid manufacturing system and captured with a Xiris XVC-1000 camera with a framerate output of 30 fps at 1280×1020 resolution. The TIG-WAAM build was deposited using a Lincoln Electric R450 Robotic Power Source, AutoDrive 4R220 Wire Drive, and a Power Wave Advanced Process Module paired with a Binzel ABITIG WH 400 W robotic GTAW torch. The camera system used was an acA800-200gc Basler ace camera at 25 fps and 800×600 resolution and equipped with a Ricoh FL-CC7528-2 M manual iris c-mount lens, a neutral density (ND) filter, and a red filter. The PAW build was deposited using an SBI PMI 400 TL3 plasma power supply paired with a WF-1 wire feeder and a TP450 plasma torch and captured using the same camera configuration as the TIG process. The Polymer build was deposited using a custom 6-axis HADDY robotic arm with pellet screw extrusion, where the irPolymer dataset was captured by four FLIR Lepton 3.5 s at 9 fps and 160×120 resolution and the visPolymer dataset was captured by an Intel RealSense D435i at 1 fps and 640×480 resolution.
Data source location	Oak Ridge National Laboratory’s Manufacturing Demonstration Facility in Oak Ridge, TN, United States
Data accessibility	Repository Name: AMVOS: Additive Manufacturing Video Object Segmentation Dataset Data identification number: doi:10.7910/DVN/5GSQTS Direct URL to data: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/5GSQTS Instructions for accessing these data: Click the Access Dataset dropdown on the dataset page and select one of the download options
Related research article	C. Wetzel, J. Haley, V. Paquit, V. Orlyanchik, H. Santos-Villalobos, Domain-Specific Adaptation: Low-Rank Adaptation Fine-tuning of SAM 2 for Manufacturing Processes. J. Intell. Manuf. (2026), accepted for publication (waiting for DOI).

1. Value of the Data

• •The data offers hand-labeled additive manufacturing video frames of additive manufacturing objects from four distinct additive manufacturing processes [1,2]. Labeled video data of these processes could not be found in public datasets due to the data’s uniqueness and the time-intensive nature of object instance hand labeling. As a result, this dataset serves as a fundamental resource for additive manufacturing VOS and provides researchers access to additive manufacturing data for VOS tasks without having to collect, label, and structure the data themselves.
• •Because this dataset follows the structure of predominant VOS datasets, such as YouTube-VOS [3], DAVIS [4], MOSE [5], and MOSEv2 [6], it can be used across VOS models such as XMem [7], XMem++ [8], SAM 2 [9], SAM 3 [10], SAMURAI [11], etc. for fine-tuning, training, or inference. It can also be used by future VOS models that continue to support this common VOS dataset structure. Furthermore, its structure permits the addition of other videos within each manufacturing process, videos of other manufacturing processes, or incorporation into larger VOS datasets.
• •The data offers specific domains in additive manufacturing that are not captured by general, internet-scale VOS datasets used for training foundation models. Consequently, this dataset serves as a resource for evaluating foundation model zero-shot performance and robustness.
• •Since each frame represents object instance segmentation, the individual frames of each dataset can be used by researchers for image-based instance segmentation models, in addition to video-based instance segmentation models.

2. Background

Traditional VOS foundation models are trained on extensive datasets that reflect the broad video distributions found on the internet [9]. These models typically achieve high accuracy and efficiency on generic video object segmentation tasks [[7], [8], [9], [10], [11]]. However, specialized datasets from domains such as additive manufacturing often fall outside these training distributions, leading to reduced segmentation accuracy when the models are applied in these settings [1,12]. Domain researchers must tolerate degraded performance, provide a custom solution, or invest in collecting, organizing, and labeling domain-specific data to adapt a VOS foundation model to their application.

This dataset helps to address this gap by introducing a dataset of raw and labeled video frames captured across multiple additive manufacturing processes. The dataset is structured to align with established video object segmentation (VOS) benchmarks, enabling straightforward adoption for AI/ML training and evaluation. This dataset served as both the primary training resource and a verification benchmark for fine-tuning and validating a domain-specific VOS model, as detailed in [1].

3. Data Description

The dataset directory structure is detailed in Fig. 1, Fig. 2 and follows the DAVIS convention [4]. As illustrated in Fig. 1(a), the root directory contains the five additive manufacturing process folders. Within each process folder are two subdirectories: Annotations and JPEGImages (Fig. 1(b)). Each of these contains train and test splits (Fig. 1(c)) that house numbered video segment directories (Fig. 1(d)). The number of video segment directories in each train and test folder reflects the 70/30 train/test split used in [1]. These folders are indexed in ascending, zero-based order within each split. Each video segment directory contains ten contiguous, zero-indexed video frames (Fig. 1(e)). Frames associated with the Annotations directory path contain palettized object instance labels in a Portable Network Graphics (.png) format. Corresponding raw frames are stored in the JPEGImages directory in a Joint Photographic Experts Group (.jpg) format.

Fig. 1.

Example path to video frames.

Fig. 2.

Dataset structure of each manufacturing process. This dataset contains five additive manufacturing datasets across four distinct processes.

Data collection for each of the reported manufacturing processes was performed using a dedicated camera system. The resolution and frame rate for each manufacturing process are summarized in Table 1.

Table 1.

Manufacturing process dataset resolutions and frame rates.

Dataset	Resolution	FPS
LHW-DED	1280 × 1020	30
TIG	800 × 600	25
PAW	800 × 600	25
visPolymer	640 × 480	1
irPolymer	160 × 120	9

Each annotated frame contains labels for the two object classes present in the video frame. In the LHW-DED, TIG, and PAW datasets, the classes are Feed Wire and Melt Pool. For visPolymer and irPolymer, the classes are Nozzle and Material. The visPolymer dataset contains visible-light recordings of the polymer process whereas irPolymer contains near-infrared recordings of the same process, captured from different viewpoints and at different time stamps. The object class nomenclature is used for readability, since the ground-truth labels are represented by RGB triplets (Fig. 3(b), (d), (f), (h), (j)). Specifically, white (255, 255, 255) represents one object (e.g., Feed Wire/Nozzle), and green (0, 255, 0) denotes the second instance (e.g., Melt Pool/Material).

Fig. 3.

Example raw images and annotations of each manufacturing process contained within the dataset [1].

4. Experimental Design, Materials and Methods

The LHW-DED dataset captures a metal deposition process in which a laser melts a continuously fed wire to generate a melt pool that solidifies layer by layer to form a 3D structure. The feed wire receives an additional energy input via resistive heating to preheat the wire prior to entering the melt pool. The data was collected on a Mazak VTC-800 G SR AM, a 5-axis deposition and 5-axis machining hybrid manufacturing platform. The images of the feed wire and melt pool were acquired using a Xiris XVC-1000 camera operated at a rate of 30 fps with a resolution of 1280 × 1020. As shown in Fig. 4(a), the camera was mounted to the extruder head of the machine with an overview perspective of the deposition process.

Fig. 4.

Manufacturing process experimental setups.

TIG-WAAM employs a feed wire and generates an electric arc between a non-consumable tungsten electrode and the welding surface to create a melt pool. The equipment used to generate the TIG dataset consisted of a Lincoln Electric R450 Robotic Power Source, AutoDrive 4R220 Wire Drive, and Power Wave Advanced Process Module. The Lincoln equipment was paired with a Binzel ABITIG WH 400 W robotic GTAW torch with the optional cold wire feeding module. The TIG torch was attached to an ABB IRB 2600 robot arm. As shown in Fig. 4(c), a Basler ace camera system, operated at 25 fps with a resolution of 800 × 600, was mounted to an external tripod to provide a side view of the deposition process.

The PAW process employs a non-consumable tungsten electrode and creates an ionized plasma stream from argon gas, which is then focused by a chilled copper nozzle. The equipment used to generate the PAW dataset consisted of an SBI PMI 400 TL3 plasma power supply paired with a WF-1 wire feeder and a TP450 plasma torch. As in the TIG setup, a Basler Ace camera system captured side view images at a rate of 25 fps and resolution of 800 × 600. The camera was mounted on an external tripod (Fig. 4(c)).

The Polymer deposition process consists of pellet screw extrusion where composite feedstock is heated, extruded through a nozzle, and fused to previously deposited layers. The extrusion and camera systems were integrated on a custom 6-axis HADDY robotic arm. The images were collected using a single Intel RealSense D435i and four FLIR Lepton 3.5 s cameras, which provided depth, visible color, and thermal measurements (up to 400 °C) of the nozzle, extruded material, and prior supporting material in a 10×10 cm region centered at the nozzle tip. The visible frames were sampled at 1 fps, and the near-infrared frames were sampled at 9 fps (Fig. 4(b), (d)).

Once the process videos were captured, each video was split into individual frames using ffmpeg [13]. Ten-frame-long, disjoint video segments were then randomly sampled and organized into video segment directories in the JPEGImages folder. Using Roboflow [14] and following the DAVIS dataset format [4], the two object instances in each video frame were manually annotated, assigned unique RGB triplet values, and saved within the Annotations directory as portable network graphics (.png) files.

Limitations

Data collection was limited to a single build per process. Consequently, the dataset contains frames from multiple videos and perspectives of the same build but does not contain video frames across builds. Furthermore, the number of frames available for annotation was limited by the frame rate of the camera system attached to the deposition process, the number of collected videos of each manufacturing process, and the general requirement that VOS datasets contain groupings of sequential video frames.

Ethics Statement

The authors have read and followed the ethical requirements for publication in Data in Brief and confirm that the current work does not involve human subjects, animal experiments, or any data collected from social media platforms.

CRediT Author Statement

Calvin Wetzel: Conceptualization, Methodology, Software, Validation, Formal Analysis, Investigation, Data Curation, Writing, Visualization; Mithulan Paramanathan: Resources, Data Curation, Methodology, Experimentation; Mario Rodriguez Parra: Resources, Data curation, Visualization; Tom Feldhausen: Resources, Investigation; Chris Masuo: Investigation, Resources, Visualization; Michael Sebok: Investigation, Data Curation, Experimentation; James Haley: Investigation, Data Curation, Experimentation, Software; Hector Santos-Villalobos: Resources, Writing – Review & Editing, Supervision, Project Administration, Funding Acquisition; Vincent Paquit: Resources, Funding Acquisition; Vladimir Orlyanchik: Resources, Project Administration, Supervision, Funding Acquisition, Writing – Review & Editing.

Data Availability

• DataverseAMVOS: Additive Manufacturing Video Object Segmentation Dataset (Original data).