lance-format/pascal-voc-2012-segmentation-lance

Pascal VOC 2012 Segmentation (Lance Format)

A Lance-formatted version of the Pascal VOC 2012 semantic segmentation split, sourced from nateraw/pascal-voc-2012. Each row pairs an inline JPEG image with the per-pixel PNG segmentation mask and a cosine-normalized OpenCLIP ViT-B-32 image embedding, so a single columnar table carries both annotation modalities and the features needed to retrieve, curate, and train against them — all available directly from the Hub at hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data.

Key features

• Inline JPEG bytes and inline PNG mask bytes in the same row — image and per-pixel segmentation travel together with no sidecar folders or mask lookups.
• Pre-computed CLIP image embeddings (image_emb, 512-dim, cosine-normalized) with a bundled IVF_PQ index for visual similarity search.
• Standard VOC class encoding — mask pixel values are class ids in 0..20 plus 255 for void, identical to the official VOC palette.
• One columnar dataset — scan image-level metadata cheaply, then fetch image or mask bytes only for the rows you actually want.

The 20 Pascal VOC foreground classes are: aeroplane, bicycle, bird, boat, bottle, bus, car, cat, chair, cow, diningtable, dog, horse, motorbike, person, pottedplant, sheep, sofa, train, tvmonitor.

Splits

Split	Rows	Notes
train.lance	1,464	Official VOC 2012 segmentation train
validation.lance	1,449	Official VOC 2012 segmentation val

Schema

Column	Type	Notes
id	int64	Row index within the split (natural join key for merges)
image	large_binary	Inline JPEG bytes
mask	large_binary	Inline PNG bytes — class id per pixel (0=background, 1-20=VOC classes, 255=void)
image_emb	fixed_size_list<float32, 512>	OpenCLIP ViT-B-32 image embedding (cosine-normalized)

Pre-built indices

• IVF_PQ on image_emb — metric=cosine

Note: the small split sizes (≤1,464 rows) sit below Lance's default partition count, so the helper falls back to a smaller num_partitions automatically. For higher recall, rebuild the index with num_partitions=16 against a local copy.

Why Lance?

1. Blazing Fast Random Access: Optimized for fetching scattered rows, making it ideal for random sampling, real-time ML serving, and interactive applications without performance degradation.
2. Native Multimodal Support: Store text, embeddings, and other data types together in a single file. Large binary objects are loaded lazily, and vectors are optimized for fast similarity search.
3. Native Index Support: Lance comes with fast, on-disk, scalable vector and FTS indexes that sit right alongside the dataset on the Hub, so you can share not only your data but also your embeddings and indexes without your users needing to recompute them.
4. Efficient Data Evolution: Add new columns and backfill data without rewriting the entire dataset. This is perfect for evolving ML features, adding new embeddings, or introducing moderation tags over time.
5. Versatile Querying: Supports combining vector similarity search, full-text search, and SQL-style filtering in a single query, accelerated by on-disk indexes.
6. Data Versioning: Every mutation commits a new version; previous versions remain intact on disk. Tags pin a snapshot by name, so retrieval systems and training runs can reproduce against an exact slice of history.

Load with datasets.load_dataset

You can load Lance datasets via the standard HuggingFace datasets interface, suitable when your pipeline already speaks Dataset / IterableDataset or you want a quick streaming sample.

import datasets

hf_ds = datasets.load_dataset("lance-format/pascal-voc-2012-segmentation-lance", split="train", streaming=True)
for row in hf_ds.take(3):
    print(row["id"], len(row["image"]), len(row["mask"]))

Load with LanceDB

LanceDB is the embedded retrieval library built on top of the Lance format (docs), and is the interface most users interact with. It wraps the dataset as a queryable table with search and filter builders, and is the entry point used by the Search, Curate, Evolve, Versioning, and Materialize-a-subset sections below.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data")
tbl = db.open_table("train")
print(len(tbl))

Load with Lance

pylance is the Python binding for the Lance format and works directly with the format's lower-level APIs. Reach for it when you want to inspect or operate on dataset internals — schema, scanner, fragments, and the list of pre-built indices.

import lance

ds = lance.dataset("hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data/train.lance")
print(ds.count_rows(), ds.schema.names)
print(ds.list_indices())

Tip — for production use, download locally first. Streaming from the Hub works for exploration, but heavy random access and ANN search are far faster against a local copy:

hf download lance-format/pascal-voc-2012-segmentation-lance --repo-type dataset --local-dir ./pascal-voc-2012-segmentation-lance

Then point Lance or LanceDB at ./pascal-voc-2012-segmentation-lance/data.

Search

The bundled IVF_PQ index on image_emb makes visual nearest-neighbour search a single call. In production you would encode a query image (or a class prototype) through OpenCLIP ViT-B-32 at runtime and pass the resulting 512-d cosine-normalized vector to tbl.search(...). The example below uses the embedding from row 42 as a runnable stand-in.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data")
tbl = db.open_table("train")

seed = (
    tbl.search()
    .select(["image_emb"])
    .limit(1)
    .offset(42)
    .to_list()[0]
)

hits = (
    tbl.search(seed["image_emb"])
    .metric("cosine")
    .select(["id"])
    .limit(10)
    .to_list()
)
for r in hits:
    print(r["id"])

Because the embeddings are cosine-normalized, metric="cosine" is the natural choice and the first hit is typically the seed row itself — a useful sanity check before tuning nprobes and refine_factor for recall.

Curate

A typical curation pass for a segmentation workflow combines visual similarity with a structural filter on the row. Stacking both inside a single filtered scan keeps the candidate set small and explicit, and the bounded .limit(200) makes it cheap to inspect before committing to anything downstream. The snippet below seeds from row 42 and restricts the candidates to rows whose mask payload is non-trivially sized — a cheap proxy for masks that actually carry foreground annotation.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data")
tbl = db.open_table("train")

seed = (
    tbl.search()
    .select(["image_emb"])
    .limit(1)
    .offset(42)
    .to_list()[0]
)

candidates = (
    tbl.search(seed["image_emb"])
    .metric("cosine")
    .where("octet_length(mask) > 2000", prefilter=True)
    .select(["id"])
    .limit(200)
    .to_list()
)
print(f"{len(candidates)} candidates")

The result is a plain list of dictionaries, ready to inspect, persist as a manifest of ids, or feed into the Evolve and Train workflows below. The image and mask columns are never read in the candidate scan, so the network traffic stays dominated by the embedding vectors rather than image or mask bytes.

Evolve

Lance stores each column independently, so a new column can be appended without rewriting the existing data. The lightest form is a SQL expression: derive the new column from columns that already exist, and Lance computes it once and persists it. The example below adds an image_bytes size and a has_mask flag, both of which can then be used directly in where clauses without recomputing the predicate on every query.

Note: Mutations require a local copy of the dataset, since the Hub mount is read-only. See the Materialize-a-subset section at the end of this card for a streaming pattern that downloads only the rows and columns you need, or use hf download to pull the full split first.

import lancedb

db = lancedb.connect("./pascal-voc-2012-segmentation-lance/data")  # local copy required for writes
tbl = db.open_table("train")

tbl.add_columns({
    "image_bytes": "octet_length(image)",
    "has_mask": "octet_length(mask) > 1024",
})

For class-level statistics — for example, a per-row list of class ids present in the mask, or a per-class pixel count — the values cannot be derived in SQL because they require decoding the PNG. Compute them once in an external table and join in by id:

import pyarrow as pa

class_stats = pa.table({
    "id": pa.array([0, 1, 2], type=pa.int64()),
    "classes_present": pa.array([[15], [7, 15], [9]], type=pa.list_(pa.int8())),
})
tbl.merge(class_stats, on="id")

The original columns and indices are untouched, so existing code that does not reference the new columns continues to work unchanged. New columns become visible to every reader as soon as the operation commits. For column values that require running a model over the image bytes (a second-pass embedding, an instance segmentation, a depth prediction), Lance also provides a batch-UDF API — see the Lance data evolution docs.

Train

Projection lets a training loop read only the columns each step actually needs. LanceDB tables expose this through Permutation.identity(tbl).select_columns([...]), which plugs straight into the standard torch.utils.data.DataLoader so prefetching, shuffling, and batching behave as in any PyTorch pipeline. For a segmentation run, project the JPEG bytes and the PNG mask bytes together; everything else, including the CLIP embeddings, stays on disk until you opt in.

import lancedb
from lancedb.permutation import Permutation
from torch.utils.data import DataLoader

db = lancedb.connect("hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data")
tbl = db.open_table("train")

train_ds = Permutation.identity(tbl).select_columns(["image", "mask"])
loader = DataLoader(train_ds, batch_size=16, shuffle=True, num_workers=4)

for batch in loader:
    # batch carries only the projected columns; image_emb stays on disk.
    # decode the JPEGs and PNGs, build (image, label) tensors, forward, backward...
    ...

Switching feature sets is a configuration change: passing ["image_emb"] to select_columns(...) on the next run skips JPEG and PNG decoding entirely and reads only the cached 512-d vectors, which is the right shape for training a lightweight linear probe over frozen CLIP features.

Versioning

Every mutation to a Lance dataset, whether it adds a column, merges class statistics, or builds an index, commits a new version. Previous versions remain intact on disk. You can list versions and inspect the history directly from the Hub copy; creating new tags requires a local copy since tags are writes.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data")
tbl = db.open_table("train")

print("Current version:", tbl.version)
print("History:", tbl.list_versions())
print("Tags:", tbl.tags.list())

Once you have a local copy, tag a version for reproducibility:

local_db = lancedb.connect("./pascal-voc-2012-segmentation-lance/data")
local_tbl = local_db.open_table("train")
local_tbl.tags.create("voc2012-clip-vitb32-v1", local_tbl.version)

A tagged version can be opened by name, or any version reopened by its number, against either the Hub copy or a local one:

tbl_v1 = db.open_table("train", version="voc2012-clip-vitb32-v1")
tbl_v5 = db.open_table("train", version=5)

Pinning supports two workflows. A retrieval system locked to voc2012-clip-vitb32-v1 keeps returning stable results while the dataset evolves in parallel — newly added class statistics or alternative embeddings do not change what the tag resolves to. A training experiment pinned to the same tag can be rerun later against the exact same image/mask pairs, so changes in metrics reflect model changes rather than data drift. Neither workflow needs shadow copies or external manifest tracking.

Materialize a subset

Reads from the Hub are lazy, so exploratory queries only transfer the columns and row groups they touch. Mutating operations (Evolve, tag creation) need a writable backing store, and a training loop benefits from a local copy with fast random access. Both can be served by a subset of the dataset rather than the full split. The pattern is to stream a filtered query through .to_batches() into a new local table; only the projected columns and matching row groups cross the wire, and the bytes never fully materialize in Python memory.

import lancedb

remote_db = lancedb.connect("hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data")
remote_tbl = remote_db.open_table("train")

batches = (
    remote_tbl.search()
    .where("octet_length(mask) > 2000")
    .select(["id", "image", "mask", "image_emb"])
    .to_batches()
)

local_db = lancedb.connect("./voc-subset")
local_db.create_table("train", batches)

The resulting ./voc-subset is a first-class LanceDB database. Every snippet in the Evolve, Train, and Versioning sections above works against it by swapping hf://datasets/lance-format/pascal-voc-2012-segmentation-lance/data for ./voc-subset.

Source & license

Converted from nateraw/pascal-voc-2012. The Pascal VOC dataset is released under its own custom license — please review before redistribution.

Citation

@misc{everingham2012pascal,
  title={The Pascal Visual Object Classes Challenge: A Retrospective},
  author={Everingham, Mark and Eslami, S. M. Ali and Van Gool, Luc and Williams, Christopher K. I. and Winn, John and Zisserman, Andrew},
  journal={International Journal of Computer Vision},
  year={2015}
}