Excavator (JCB) Image Dataset - Construction Vehicle Photographs

The Excavator (JCB) Image Dataset is a computer-vision photograph collection built by DataCluster Labs: more than 3,000 excavator photographs gathered from over 200 distinct locations worldwide by roughly 1,000 crowdsource contributors shooting on ordinary mobile phones between 2020 and 2022. Every frame lands at 1920x1080 resolution or above, and the set deliberately spans day, night and varied distances and viewpoints - the publisher describes it as an intentionally challenging corpus for vehicle detection on construction sites, which is exactly why it is worth having.

Inside construction machinery and heavy transportation equipment data, this record occupies the imagery layer. Where listing feeds give you price, hours and model year per machine, this set gives you pixels per machine class - excavators as they actually appear on working sites, backlit, half-dusty, parked at dusk, framed too far away. That is the distribution your detector meets in production, and training against anything cleaner produces models that flatter themselves. Get a sample of this dataset and see whether the frames match your deployment conditions.

One record per photograph. The row below is the verified example observed during cataloging, unfolded into the delivered column order:

image_file   : 20221124_21_59_30_129_000_wJlYqohWF6asvZqLrt8adDivdkw1_T_3000_4000.jpg
captured_at  : 2022-11-24 21:59:30.129     <- parsed from the leading filename stamp
device_token : wJlYqohWF6asvZqLrt8adDivdkw1  <- camera-source string, kept verbatim
tail_tokens  : _T_3000_4000                  <- trailing sequence tokens
subject      : one excavator, one full frame
resolution   : 1920x1080 or above
annotations  : bounding boxes, COCO / YOLO / PASCAL-VOC / TFRecord layout

Read what the row already says. The filename is not decoration - it encodes the capture timestamp and camera-source metadata, so captured_at resolves without opening the file or trusting a sidecar, and the 21:59 stamp on the very first observed frame is a live demonstration of the night-time half of the day/night spread. The trailing token block carries no documented meaning, so it travels verbatim rather than re-labelled with a guess. Request a sample and every row arrives with the same seven slots filled, annotations in whichever of the four layouts your stack reads natively.

Two documented fields define the core record, verified during cataloging against the collection itself. image_file carries the photograph itself - one flat .jpg per excavator frame, with the capture timestamp and device metadata riding in the filename rather than in separate metadata columns. annotation_layout carries the bounding-box label set in one of the four interchange layouts - COCO, YOLO, PASCAL-VOC or TensorFlow Record - selected per your request rather than imposed.

Field definitions here were inferred during cataloging rather than published as a formal data dictionary by the compiler, which is why the dictionary below states definitions and examples plainly and pushes nothing speculative into the typed core. What sits outside the verified two is listed under additional fields on request rather than promised blind.

Geography - over 200 distinct crowdsourced locations worldwide, contributed by roughly 1,000 photographers. One honest limit: no coordinates attach to individual images, so geography is a property of the collection's spread, not a joinable column. If your model needs spatial context, pair the frames with the tabular registration and pricing sets in this catalog rather than expecting lat/long here.

Temporal - every photograph was captured between 2020 and 2022, which makes the corpus a fixed benchmark rather than a moving feed. That fixity is a feature: accuracy measured on it in March equals accuracy measured on it in November, so comparisons across models, papers and quarters stay honest. Delivery cadence remains yours to set - daily, weekly, or hourly - for how the same records reach your environment.

Granularity - one record per photograph, more than 3,000 records in total, each at 1920x1080 or above. No aggregation layers sit between you and the frames: no category rollups, no monthly composites, nothing smoothed. Roughly 798 MB of imagery in total, carried as flat .jpg files keyed by their capture-timestamp filenames.

API, files, or your warehouse. Daily, weekly, or hourly.

Name the frame volume, the annotation layout and the split you want when you request the sample - the full 3,000-plus corpus with boxes in TensorFlow Record for a training run, or a few hundred frames heavy on night captures for a low-light evaluation. The sample ships first either way, and the ongoing delivery lands shaped identically, so the loader you prototype against the sample survives into production unchanged. Field dictionary, sample rows and a coverage profile travel alongside every delivery.

• Construction vehicle detection - train and benchmark excavator detectors against a corpus assembled to be hard: night frames, long distances, awkward viewpoints, real dust and shadow instead of stock-photo lighting.
• Low-light and distance robustness testing - the deliberate day/night/distance spread turns "our model degrades after dark" from an anecdote into a measurable recall delta with a fixed test bed behind it.
• Site-safety and proximity analytics - detect and localize excavators in yard and site footage as the first stage of proximity-warning and exclusion-zone systems before narrower classes are layered on.
• Equipment marketplace visual features - prototype visual search and automatic photo tagging for rental and dealer platforms, where listing photos need to be recognized and classified at upload time.
• Insurance and claims triage - pre-screen submitted damage photography for machine type and framing before an adjuster ever opens the file.
• Academic teaching and coursework - a domain-specific counterpart to generic benchmarks, small enough for a semester project and messy enough to teach why product image datasets differ from laboratory ones.

Data scientists and ML engineers get a fixed, challenging test bed whose population never shifts underneath a benchmark run - see construction machinery data for data scientists. Developers and builders smoke-test detection services end to end against imagery that resembles production input rather than curated samples - construction machinery data for developers and builders. Competitive intelligence and product teams evaluate whether a visual-search or auto-tagging feature is worth building before committing to a collection pipeline of their own. Journalists, academics and students teach applied computer vision on construction-domain imagery whose subject matter everyone in the room recognizes.

Provenance note - compiled by DataCluster Labs from roughly 1,000 crowdsource contributors using consumer mobile phones across more than 200 locations. Consumer capture is the point: focal lengths, sensor noise and framing vary the way they vary in the wild, which laboratory rigs never reproduce.

Methodology note - field definitions were inferred during cataloging from the collection itself rather than read off a published dictionary, and the exact count beyond the 3,000-plus floor went unpublished because the full enumeration exceeded what verification could complete. Sizing questions get answered inside your sample, not asserted here.

Completeness note - Datadory scores this record 6/10: strong on scale, resolution and intentional difficulty, held back by the absent per-image coordinates, the unpublished exact count and an annotation layer that requires confirmation per cut. None of that is hidden - it is priced into the score.

Where to go next - inside Construction Machinery & Heavy Transportation Equipment, the tabular complement is Heavy Equipment Pricing Data (US, Kaggle): 1,742 US listings with price, manufacturer, hours and region - pixels there, prices here. MachineryTrader Construction Equipment Listings adds the live multi-category supply view, and Ritchie Bros. Heavy Equipment Auction Data contributes realized transaction values. See where every record ranks on the best construction machinery heavy transportation equipment datasets list, browse the construction machinery heavy transportation equipment data hub, or start from the construction machinery heavy transportation equipment data guide. Background reading: benchmark dataset and the Kaggle Datasets Hub glossary entries, plus the Kaggle source profile.