Places

Places, also called the Places Database or MIT Places, is a collection of photographs of places — bedrooms, coasts, airport terminals, forest paths, zen gardens — each labeled with the type of environment it shows. Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba and Aude Oliva, from MIT, Princeton University and the Universitat Oberta de Catalunya, introduced it at NIPS 2014 with a 205-category benchmark, Places205, and a convolutional network trained on it. The second generation, Places2, was released for a challenge held with ILSVRC in 2015, and its 365-category subsets Places365-Standard and Places365-Challenge appeared in 2016; the full database of 10,624,928 images in 434 categories is described in the authors’ 2018 TPAMI paper. Where ImageNet is object-centric, its images labeled by the object they contain, Places is scene-centric, and the networks trained on it (Places-CNNs) became the standard starting point for scene recognition, a form of image classification.

Purpose

Zhou et al. (2014) started from a gap. Convolutional networks trained on ImageNet had transformed object recognition, yet their features scored on scene benchmarks only in the range of hand-designed features. The authors argued that object-centric training was the reason: iconic photographs of objects lack the variety of layouts and contents that pictures of whole environments carry. The existing scene datasets were too small to train a network: Scene15 had 15 categories with a few hundred images each, MIT Indoor67 had 67 indoor categories, and the SUN database’s 397 well-sampled categories had at least 100 images each. Places was built to be large enough to train a CNN on scenes alone, 60 times larger than SUN by the 2014 paper’s count.

The project page states the second aim in terms of human cognition: categories are defined by function at the entry level of an environment, so that a home bedroom, a hotel bedroom and a nursery are different classes because people act differently in them (official page, as of October 2026). The TPAMI paper frames the database as a quasi-exhaustive list of the kinds of environments in the world, bounded below by spaces a human body would fit into, such as a closet or a shower, and estimates that its 434 categories cover about 98 percent of the places a person encounters; the paper does not say how that percentage was computed.

Same architecture, different data, different features: Zhou et al. (2014) trained the Caffe reference network, an AlexNet-style network, on Places205 and compared it unit by unit with the same network trained on ImageNet. From the second pooling layer onward, the averages of the images that most activated each unit looked increasingly like object blobs for the ImageNet network and like landscapes with spatial structure for the Places network (§4.1). The TPAMI paper (§5.5) segmented the images that most activate each conv5 unit and found more units for object parts, such as dog and human heads, in the ImageNet network and more for scene parts, such as beds, chairs and buildings, in the Places network. A companion ICLR 2015 paper showed that object detectors emerge inside a CNN trained only to classify scenes, without any object labels (Zhou et al., 2015).

Dataset Size

The database grew in stages, and each paper reports its own counts.

Stage	Categories	Images	Source
Places, after two annotation rounds (2014)	476	7,076,580	NIPS 2014 §2.1; TPAMI §2.2.2
Places205 training set	205	2,448,873	NIPS 2014 §4; Places205 page
Places2, after four annotation rounds	434	10,624,928	TPAMI §1, Fig. 2, §2.2.4
Places365-Challenge training set	365	8,026,628	development kit README
Places-Extra69	69	98,721 training, 6,600 test	download page (archived 2021); development kit README

The Places2 home page rounds the database to more than 10 million images in “400+” categories, with 5,000 to 30,000 training images per class; the Places365-Challenge training set, by the development kit’s count, reaches 40,000 per class (as of October 2026). Places365-Standard and Places-Extra69 together hold the 434 categories (365 + 69).

Images are colour photographs of at least 200 × 200 pixels as downloaded. The Places365 archives come in two forms: high-resolution images resized so that the shorter side is 512 pixels (smaller originals left unchanged), and $256 \times 256$ images resized without preserving the aspect ratio (download page, archived 2021).

Classes

From SUN. Places inherited its category list from the SUN (Scene UNderstanding) database of Xiao et al. (2010). To make that list, the SUN team took the roughly 70,000 WordNet terms used by the Tiny Images dataset and kept every concrete noun that could complete “I am in a place” or “Let’s go to the place”. They excluded proper names (Grand Canyon), terms without a specific visual identity (territory, workplace), non-navigable scenes such as a desktop, vehicles except views of their interiors, and mature content, but kept building types such as skyscraper and hangar because buildings activate scene-processing areas of the brain. About 2,500 initial terms became 899 categories after synonyms were bundled and scenes sharing a word but differing visually were split, for instance indoor and outdoor views of a church. As a completeness check, seven people noted their surroundings every half hour for a total of 284 hours; all 52 scene categories they reported were in the list (Xiao et al., 2010, §2 and footnote 1).

Refined for Places. In the last construction step, Zhou et al. (2018, §2.2.4) found that SUN’s list still contained near-synonyms such as “ski lodge” and “ski resort” or “garbage dump” and “landfill”, and merged 46 such pairs. They also identified 53 pairs of categories whose images workers confused, such as “canyon” and “mountain”, “pond” and “lake”, or “operating room” and “hospital room”, and re-annotated them in a dedicated round (see Annotation Types).

Subsets by size. Each benchmark takes the categories with enough images. Places205 has the 205 categories with more than 5,000 images in the 2014 collection. Places365 has the 365 of the 434 Places2 categories with more than 4,000 images each, and the other 69 were released as Places-Extra69 (TPAMI §3; development kit). Places88 has the 88 categories that Places, SUN and ImageNet share with at least 1,000 ImageNet images, used to compare the three datasets. The class lists and names changed between versions: only 182 categories of Places205 map directly to Places365, because some Places205 categories were merged (TPAMI §5.1).

Names. Places365 category names keep SUN’s slash notation, such as /b/bedroom, with qualifiers for categories split by viewpoint or type: /s/swimming_pool/indoor and /s/swimming_pool/outdoor, /c/canal/natural and /c/canal/urban (categories_places365.txt). An accompanying file labels 161 of the 365 categories indoor and 204 outdoor (our count of IO_places365.txt at commit 8a953ed); the TPAMI paper (Fig. 1) groups the categories into three macro-classes, indoor, nature and urban, and the project publishes a scene hierarchy as a spreadsheet.

Annotation Types

Every image has one scene label, assigned in up to four rounds of crowdsourced annotation on Amazon Mechanical Turk (AMT). The two papers describe the same process, with more detail in the TPAMI version (§2.2).

1. Query and download. Each category name was combined with each of 696 common English adjectives, such as messy, spare, sunny or desolate, and sent to Google Images, Bing Images and Flickr. The adjectives raised both the number of images available per category and the variety of their appearance (“stylish kitchen”, “messy kitchen”, “wooded kitchen”); the list is in the Places365 repository (adjectives_download.csv, 696 entries). Duplicate URLs were removed, and PCA-based near-duplicate removal ran within each category and against the same category in SUN, so that Places and SUN share no images and can be combined. The 2014 paper reports more than 40 million images downloaded; the TPAMI paper reports about 60 million images with unique URLs.
2. Two rounds of yes/no verification. Workers saw a category’s definition, examples of true and false images, and then one image at a time, answering whether it belongs to the category. Because over half of downloaded images had not been true exemplars when SUN was built, the first round defaulted to NO, so a worker had to act to accept an image. Images accepted in round one went to a second round that defaulted to YES, so a worker had to act to reject; it relabeled 25.4% of them as NO. Each task (a HIT) held 750 images plus 30 positive and 30 negative control images with SUN ground truth, and a worker’s HIT was kept only with at least 90% accuracy on the controls. A third iteration was tested on a few exemplars and abandoned because it changed few labels. These two rounds produced the 7,076,580 images in 476 categories.
3. Classifier-assisted round. About 53 million downloaded images remained unassigned to any of the 476 categories, though many belonged to one (a bedroom found while searching for living rooms and marked negative). An AlexNet trained on 1,000 images from each of the 413 categories that had at least that many, reaching 32% accuracy on a 50-per-category validation set, classified those images; for categories with too few exemplars, images it predicted with confidence above 0.8 went back to AMT with the default-NO interface. This added about 3 million images and flattened the category sizes: 401 categories passed 5,000 images and 240 passed 20,000.
4. Separating similar classes. For the 53 confusable pairs, a new interface showed typical examples of both categories and asked workers to assign each image to one, the other, or neither, a question the authors note is easier than a yes/no judgment at a blurry boundary. After this round the database held 10,624,928 images in 434 categories.

Places has no bounding boxes, object labels or segmentations. The project’s web demo also predicts SUN scene attributes and indoor or outdoor type; the Places205 page offers 102 attribute detectors built on Places205-AlexNet features and per-category indoor/outdoor labels for voting over the top five predictions. These are outputs of models, not annotations (Places205 page and Places365 README, as of October 2026).

Comparing Datasets: Density and Diversity

The 2014 paper proposed two measures for comparing image collections that need only judgments of which of two pairs of images is more alike, never an absolute similarity. A good dataset, the authors argued, should be dense, its images having close neighbours, and diverse, with a high variability of appearance and viewpoint. A set of 100,000 photographs of one bedroom would be very dense and not at all diverse.

For image sets $A_1, \dots, A_N$ and an image distance $d$, the relative density of $A_1$ is

$$\mathrm{Den}_{A_2,\dots,A_N}(A_1) = p\left(d(a_{11}, a_{12}) < \min_{i=2,\dots,N} d(a_{i1}, a_{i2})\right),$$

where each $a_{i1}$ is a random image of $A_i$ and $a_{i2}$ a near neighbour of it in the same set. The relative diversity uses two random images from each set instead:

$$\mathrm{Div}_{A_2,\dots,A_N}(A_1) = 1 - p\left(d(a_{11}, a_{12}) < \min_{i=2,\dots,N} d(a_{i1}, a_{i2})\right).$$

For two sets, $A$ is denser than $B$ if and only if $\mathrm{Den}_B(A) > \mathrm{Den}_A(B)$, and likewise for diversity (NIPS 2014, Eqs. 1–2). The diversity measure was inspired by Simpson’s index, which gives the probability that two individuals drawn from an ecosystem belong to the same species. Images have no sub-category labels to play the part of species, so the measure compares datasets instead.

In the experiments, people played the role of $d$. On each AMT trial a worker saw four pairs from each of Places88, SUN88 and ImageNet88, twelve pairs in all, and picked the most similar pair; there were 40 trials per category with two workers each, over the 88 shared categories. For density, each pair was a random image and its fifth nearest neighbour by Euclidean distance between Gist descriptors, skipping the first four neighbours to avoid near-duplicates, with 25 trials per category. Pairs from SUN were chosen as the most similar 50% of the time, ImageNet pairs 33% (a figure only the TPAMI version states) and Places pairs 17%, giving average relative diversities of 0.50, 0.67 and 0.83. The three datasets were, on average, very similar in density (NIPS 2014, §3.2). By our reading, a trial with four pairs per dataset estimates a related but not identical probability to the one-pair definition above. The TPAMI paper (§4.1) repeats the diversity experiment and its numbers but drops the density measure and its results.

The code below computes both measures for synthetic image sets with known variety. A renderer draws $32 \times 32$ outdoor scenes (sky, ground, a horizon with camera roll, buildings and a sun) from wide parameter ranges (the varied set) or from narrow ranges around one layout (the prototypical set). A third set imitates the single-bedroom thought experiment: one fixed scene with the camera shifted by at most a pixel. A Euclidean distance between $8 \times 8$ block means stands in for the human judgment, and each probability is estimated from 20,000 Monte Carlo trials.

import numpy as np

S = 32  # image side in pixels

def render(rng, varied):
    """A 32x32 grayscale outdoor scene: sky, ground, buildings and a sun.
    varied=True draws the layout from wide ranges; varied=False keeps every
    parameter near one prototypical layout (two buildings, sun top right)."""
    y, x = np.mgrid[0:S, 0:S]
    w = 1.0 if varied else 0.4                       # width of the parameter ranges
    horizon = 20 + w * rng.uniform(-10, 8)           # row of the horizon
    tilt = w * rng.uniform(-0.4, 0.4)                # camera roll, as horizon slope
    sky = 0.75 + w * rng.uniform(-0.3, 0.2)
    img = np.where(y < horizon + tilt * (x - S / 2), sky - 0.01 * y, 0.35)
    for b in range(rng.integers(0, 5) if varied else 2):
        if varied:
            cx, bw, bh = rng.uniform(0, S), rng.uniform(3, 10), rng.uniform(4, 16)
        else:
            cx, bw, bh = rng.normal(10 + 12 * b, 2), rng.uniform(5, 7), rng.uniform(8, 12)
        inside = (abs(x - cx) < bw / 2) & (y > horizon - bh) & (y < horizon)
        img = np.where(inside, 0.15 + 0.3 * w * rng.random(), img)
    if not varied or rng.random() < 0.5:
        sx, sy = (rng.uniform(4, 28), rng.uniform(3, 10)) if varied else (25, 6)
        img = np.where((x - sx) ** 2 + (y - sy) ** 2 < 6, 1.0, img)
    return img + rng.normal(0, 0.02, img.shape)      # sensor noise

def one_room(rng):
    """The paper's thought experiment: every picture taken in the same place;
    only the camera position jitters, by up to one pixel."""
    base = render(np.random.default_rng(0), varied=False)
    dy, dx = rng.integers(-1, 2, size=2)
    return np.roll(base, (dy, dx), axis=(0, 1)) + rng.normal(0, 0.02, base.shape)

def distances(images):
    """Euclidean distances between 8x8 block means (tiny-image descriptors)."""
    f = np.array([im.reshape(8, 4, 8, 4).mean(axis=(1, 3)).ravel() for im in images])
    sq = (f ** 2).sum(1)
    return np.sqrt(np.maximum(sq[:, None] + sq[None, :] - 2 * f @ f.T, 0))

def relative(sets, rng, kind, trials=20000, k=5):
    """Monte Carlo estimate of the relative diversity (Eq. 2) or density (Eq. 1)
    of each set with respect to all the others. In every trial each set offers
    one pair; the set whose pair is closest wins the trial. Diversity pairs are
    two random images; density pairs are a random image and its k-th nearest
    neighbour in the same set (the paper used the 5th)."""
    closest = []
    for images in sets:
        d = distances(images)
        n = len(d)
        i = rng.integers(0, n, trials)
        if kind == "diversity":
            j = (i + rng.integers(1, n, trials)) % n     # a second, different image
            closest.append(d[i, j])
        else:
            closest.append(np.sort(d, axis=1)[i, k])     # column 0 is the image itself
    p = np.bincount(np.argmin(closest, axis=0), minlength=len(sets)) / trials
    return 1 - p if kind == "diversity" else p

# A varied set of n images against a prototypical set of 300, five seeds each.
for n in (300, 1200, 4800):
    div, den = [], []
    for seed in range(5):
        rng = np.random.default_rng(seed)
        sets = [[render(rng, True) for _ in range(n)],
                [render(rng, False) for _ in range(300)]]
        div.append(relative(sets, rng, "diversity")[0])
        den.append(relative(sets, rng, "density")[0])
    print(f"varied {n:>4} vs prototypical 300: Div {min(div):.2f}-{max(div):.2f}, "
          f"Den {min(den):.2f}-{max(den):.2f}")

# Add the thought experiment: 300 pictures of one place.
rng = np.random.default_rng(0)
sets = [[render(rng, True) for _ in range(300)],
        [render(rng, False) for _ in range(300)],
        [one_room(rng) for _ in range(300)]]
div, den = relative(sets, rng, "diversity"), relative(sets, rng, "density")
for name, a, b in zip(["varied", "prototypical", "one room"], div, den):
    print(f"{name:>12}: Div {a:.3f}  Den {b:.3f}")
print(f"sums over the three sets: Div {div.sum():.3f}, Den {den.sum():.3f}")

Output:

varied  300 vs prototypical 300: Div 0.77-0.80, Den 0.23-0.27
varied 1200 vs prototypical 300: Div 0.78-0.81, Den 0.33-0.40
varied 4800 vs prototypical 300: Div 0.78-0.80, Den 0.47-0.50
      varied: Div 0.979  Den 0.000
prototypical: Div 0.947  Den 0.000
    one room: Div 0.074  Den 1.000
sums over the three sets: Div 2.000, Den 1.000

With equal sizes, the varied set is clearly more diverse and clearly less dense than the prototypical one. Growing it sixteenfold leaves its diversity where it was, because two random images from a wide distribution stay far apart however many there are, but raises its density to parity, because a larger sample puts each image’s fifth neighbour closer. This is the pattern Zhou et al. reported for Places: as dense as SUN and ImageNet, and more diverse, with far more images per category. The single-place set wins every density trial and almost no diversity trial, as the thought experiment predicts. Because exactly one set wins each trial (ties have probability zero here), the densities of $N$ sets always sum to 1 and their diversities to $N - 1$; this is why the paper’s three diversities, 0.50, 0.67 and 0.83, sum to 2 and why identical datasets would each score $2/3$. The measures are relative: they rank sets under one distance, and a different $d$, such as human judgment, can rank them differently.

Splits

Benchmark	Classes	Training	Validation	Test	Source
Places205 (2014)	205	2,448,873 (5,000–15,000 per class)	100 per class	200 per class (41,000)	NIPS 2014 §4; TPAMI §3
Places2 Challenge 2015	401	8.1 million (4,000–30,000 per class)	about 20,000	about 381,000	challenge page, archived 2015
Places365-Standard (2016)	365	1,803,460 (3,068–5,000 per class)	36,500 (100 per class); 50 per class in TPAMI §3 and on the download page	328,500 (900 per class)	development kit README; TPAMI §3
Places365-Challenge (2016)	365	8,026,628 (3,068–40,000 per class)	as Standard	as Standard	development kit README
Places-Extra69	69	98,721	none	6,600 (100 per class where available)	development kit README; download page

Places365-Standard caps each category at 5,000 training images, and the first 5,000 images (or all, if fewer) of each Places365-Challenge category are the Standard ones. The training, validation and test sets do not overlap. The sources disagree on two counts. For the Places365 validation set, the development kit’s README and file naming give 36,500 images, 100 per class, while the TPAMI paper and the download page say 50 per class; the 2016 challenge page’s “36K images for validation” agrees with the development kit, but we could not settle this without the files. For the images that Places365-Challenge adds, the README and download page say about 6.2 million, which matches the difference of the README’s two training counts (6,223,168), while the TPAMI paper’s §5.3 says 5.5 million. Places-Extra69 holds back 100 test images only in categories that have enough, which is why its test set has 6,600 images rather than 6,900.

Places88 is used only for the dataset comparisons above. It contains just the Places images from the second annotation round of the 2014 collection, and its ImageNet88 and SUN88 counterparts have no images in common with it (TPAMI §3).

Tasks

• Scene classification: predict the scene category of a photograph, the task of every Places benchmark and challenge. Zhou et al. (2018, §1) present a place category as a compact substitute for listing the objects in a picture and their spatial relations, much as an object name replaces a list of its parts.
• Scene-centric feature learning: train a network on Places and reuse its features for other scene tasks (transfer learning), the use that motivated the 2014 paper.

The Places team also ran challenges on other data. The Scene Parsing Challenge 2016, held alongside the Places Challenge, used the pixel annotations of ADE20K, not Places scene labels, and so did the Places Challenge 2017. That challenge announced three tasks (scene parsing, scene instance segmentation and semantic boundary detection), but its results page lists results only for scene parsing and instance segmentation (official challenge and results pages).

Evaluation

Places benchmarks use the flat top-5 error of ILSVRC classification, described in the ImageNet article; accuracy covers top-$k$ scoring in general. The Places challenge pages define it so that it could accept several ground-truth labels per image: an algorithm returns labels $l_1, \dots, l_5$, the image has ground-truth labels $g_1, \dots, g_n$, and the image’s error is

$$e = \frac{1}{n} \sum_{k=1}^{n} \min_{j} d(l_j, g_k), \qquad d(x, y) = \begin{cases} 0 & x = y \\ 1 & \text{otherwise,} \end{cases}$$

averaged over all test images. In the 2015 and 2016 editions $n = 1$, one label per image, so $e$ is 0 if the true category is among the five guesses and 1 otherwise. The organizers motivate the five guesses by scenes that fit several names: a bar can also be a restaurant, and people call one place a forest path, a forest or woods.

The development kit’s MATLAB routine eval_cls.m (commit 49a959d, read line by line) implements this. A submission is a text file with one line per test image, in file-name order, each giving exactly five category indices from 0 to 364 ranked by confidence. The routine matches predictions to the ground-truth list by file name, rejects duplicate file names, and returns the error for the first one to five guesses; only the five-guess error decides a challenge. Test labels of Places365 are not released. Results are therefore reported either on the validation set or through the official evaluation server, which, like the Places205 leaderboard, did not respond with a working page when checked in October 2026 (see Resources).

Two conventions affect published numbers. The TPAMI baselines average class scores over ten crops of each test image (Tables 1–2); its Table 1 gives the Places205-AlexNet 67.52% top-1 accuracy on SUN205, where the 2014 paper, which does not mention crops, gave the Places-CNN 66.2%. And Zhou et al. (2018, §5.1) call top-5 accuracy the more suitable criterion and top-1 accuracy possibly ill-defined for many scene categories, because one environment fits several labels.

License

The data come under separate statements for images, annotations, models and code (as of October 2026).

• Annotations. The Places205 page states that “the annotation can be used under the Creative Common License (Attribution CC BY)”, without naming a version of the license. The Places2 pages state no separate annotation license.
• Images. The Places205 page and the Places365 README both state that the copyright of all images belongs to the image owners. The terms on the Places2 download page (archived December 2021 copy) restrict use to non-commercial research and educational purposes, forbid redistributing the images, disclaim warranties by the Massachusetts Institute of Technology (including non-infringement), and require users to accept full responsibility for their use and to indemnify MIT, including against claims over copies of copyrighted images.
• Access. Both home pages say the database and CNNs are provided “for academic research and education purposes”. In October 2022 the download page said the dataset was under maintenance and asked users to email the maintainer (archived copy of 2 October 2022). As of October 2026, it says a version 2 of the database is in preparation and that anyone who needs the legacy data, the original Places365 or Places205, for research should fill in a request form; the Places205 page asks users to register. We did not register or fill in the form, so we cannot say which terms the form now attaches.
• Models and code. The Places365 README states that the pretrained Places-CNN models can be used under the same “Creative Common License (Attribution CC BY)” with credit to the paper or the project page. The repository’s code carries the MIT License, and the development kit an MIT-style permission notice (COPYRIGHT, “Copyright (c) 2015 Places365”).

How the CC BY statement for annotations sits with the non-commercial terms for the images is not addressed on the official pages, and the terms are unclear on that point. In the released files the labels are folder names and entries in image lists, so in practice the two travel together (our reading).

Limitations and Bias

• One label for a scene with several. Each image has a single category, though environments are inherently multi-label (TPAMI §5.1). Among the validation errors of the Places365 VGG network, the authors identify two kinds: unusual activities in a scene, such as a group photograph at a construction site or camping in a junkyard, and images composed of several scene parts that one label cannot describe (Fig. 11). Top-5 scoring absorbs some of this by design, at the cost of rewarding a correct label anywhere among five guesses.
• Blurry category boundaries. Field and forest grade into each other (Fig. 5), and workers mixed up 53 category pairs before the fourth annotation round. Category definitions also changed between versions: Places365 merged some Places205 categories (TPAMI §5.1).
• Label quality is not measured. Only HITs with at least 90% accuracy on the control images were kept, and the second round overturned 25.4% of first-round positives; neither paper reports an audit of the final labels’ accuracy. In the third round, an AlexNet with 32% validation accuracy chose which of the 53 million images reached workers. By our reasoning, images the classifier did not rank highly were never offered for some categories, so the added images favour what that network already recognized.
• Web search as the source. Images are those that three search engines return for a category name and an adjective, filtered to colour images of at least 200 × 200 pixels. Category sizes follow availability on the web — the TPAMI paper’s example is many more bedrooms than caves — which the Places2 home page describes as consistent with real-world frequencies of occurrence; Places365-Standard caps them at 5,000 per class. The adjectives widen appearance within a category but cannot add places the search engines do not index.
• Cross-dataset generalization. Training linear SVMs on ImageNet-CNN features of Places88, SUN88 and ImageNet88 and testing on each, Zhou et al. (2014, Fig. 4) found the best accuracy for a fixed number of training images always on the training set’s own test set, the drop that Torralba and Efros (2011) attribute to dataset bias. With all its training images, Places88 gave the best result on two of the three test sets. In the legends, which we read as the accuracies with all training images, SUN88 was classified best from Places88 (69.5% against 63.3% from SUN88 itself), while ImageNet88 was classified best from ImageNet88 (65.6% against 60.3% from Places88).
• People. As of October 2026, neither paper nor the official pages say anything about people who appear in the images, consent, faces or requests for removal. In their audit of large image datasets, Birhane and Prabhu (2021, Table 1) list Places among datasets containing images of people, with zero images collected with consent; dataset bias discusses that audit.
• Availability. The original download links have been replaced by a request form, and the evaluation server and the Places205 leaderboard did not respond with working pages in October 2026. We infer that new results on the Places365 test set cannot currently be obtained, which leaves the validation set as the de facto test set.

Common Uses

• Benchmark for scene classification. Places365-Standard, usually its validation set, is the common test of scene recognition, and papers on general architectures report it next to ImageNet; the squeeze-and-excitation papers, for example, report scene classification on Places365.
• Pretrained scene networks. The project released Places-CNNs for Caffe, Torch and PyTorch: AlexNet, GoogLeNet, VGG-16, ResNet-152 fine-tuned from ImageNet, ResNet-152 and a pre-activation ResNet-50 trained from scratch, ResNet-18 and DenseNet-161, and hybrid networks trained on the 1,000 ImageNet classes plus the 365 Places classes (Places365 README, commit 8a953ed). They supply features or initial weights for scene recognition, scene attributes and indoor/outdoor prediction.
• Interpretability. Places-CNNs have been used to study what individual units learn, starting with the emergence of object detectors in a scene network (Zhou et al., 2015). The project’s demo shows class activation maps (Zhou et al., 2016) for its scene predictions (Places205 page).
• Training data for image synthesis. Places images are a common training and test set for image inpainting; LaMa and latent diffusion models, for example, report results on $512 \times 512$ Places crops.

Historical Milestones

Scene classification results on Places, with their source and protocol. All Places accuracies are on the official test sets unless stated.

Year	Result	Protocol	Source
2014	Places205-CNN (Caffe reference network, about six days on one Tesla K40): 50.0% top-1, 18.9% top-5 error on Places205; ImageNet-CNN features with a linear SVM: 40.8%	Places205 test, single label	NIPS 2014, Table 1 and §4.2
2014	Places-CNN fc7 features with a linear SVM: 54.32% on SUN397 (50 training images per class) and 68.24% on MIT Indoor67, against 42.61% and 56.79% for ImageNet-CNN features; the previous bests the authors cite were 47.20% and 66.87%	Standard splits of each benchmark	NIPS 2014, Table 2 and supplement
2015–16	Places205-GoogLeNet 55.50% / 85.66% (on the leaderboard from May 2015) and Places205-VGG 58.90% / 87.70% top-1 / top-5	Places205 test, ten crops	TPAMI, Table 1; leaderboard (archived 2022)
2015	Places2 Challenge at ILSVRC 2015 (ICCV): 401 classes; winner WM (Li Shen, Zhouchen Lin), VGG-like networks, 16.87% top-5 error	Challenge test set, provided training data	2015 challenge and results pages
2016	Places205 leaderboard: SamExynos 64.10% / 90.65% top-1 / top-5, submitted 17 April 2016 and still first in the September 2022 archive	Places205 test server	leaderboard (archived 2022); TPAMI, Table 1
2016	Places365-Standard baselines: Places365-VGG 55.19% / 85.01%, Places365-ResNet (ResNet-152 fine-tuned) 54.65% / 85.07% top-1 / top-5	Places365 test, ten crops	TPAMI, Table 2
2016	Places Challenge at ILSVRC 2016 (ECCV): 92 valid submissions from 27 teams; Hikvision 9.01%, MW 10.19%, Trimps-Soushen 10.30% top-5 error, against 14.9% for the best Places365-Standard baseline	Places365 test, Places365-Challenge training data only (networks pretrained on ImageNet or earlier Places allowed), at most five submissions per team; the results page describes the third-place entry as “With extra data.”	2016 results page; TPAMI §5.3 and Fig. 13
2017	Places365-VGG fc7 features with a linear SVM: 63.24% on SUN397; the hybrid ImageNet + Places365 VGG gave the best average over eight scene and object benchmarks	Same benchmark protocols as 2014, SUN397 over 5 splits	TPAMI, Table 3

The 2014 comparison set the pattern for what followed: on scene benchmarks the Places features won, on object benchmarks (Caltech-101, Caltech-256, Stanford Action40) the ImageNet features won by 12 to 22 points. A Hybrid-CNN trained on both datasets (3.5 million images in 1,183 categories) came within 2.5 points of the better network on every benchmark and beat both on four (NIPS 2014, Table 3). In the TPAMI paper’s own comparison (§5.3), the winning 2016 entry beat the best baseline by about six points of top-5 accuracy, and the authors point out that the entries trained on the much larger Places365-Challenge set while the baselines used Places365-Standard; they add that their ResNet baseline was fine-tuned rather than trained from scratch and may be under-trained (footnote 5).

Places-CNNs predicting more classes did not lose accuracy on the old ones: on the 182 categories Places205 and Places365 share, the Places365-AlexNet reached 57.7% top-1 on Places365 validation images against 57.2% for the Places205-AlexNet (TPAMI §5.1). In the wild, the web demo’s users rated its top-5 predictions correct about 72% of the time, over 9,925 pieces of feedback from October 2014 to May 2016 (TPAMI §5.2).

Related Datasets

• SUN database (Xiao et al., 2010): 899 scene categories and 130,519 images, the source of Places’ category list and of the control images in its annotation; its 397-category benchmark, SUN397, remains a standard test for scene features, and the SUN Attribute database provides the scene attributes the Places demo predicts.
• Scene15 and MIT Indoor67: the earlier scene benchmarks, with 15 categories and 67 indoor categories.
• ImageNet: the object-centric counterpart; it contains scene categories only in small proportion, 128 of the 205 Places205 categories (NIPS 2014, §2). Places88 compares the two directly.
• ADE20K: densely annotated scenes with objects and parts, from the same MIT group, used for the scene parsing challenges of 2016 and 2017.

Resources

• Places2 project page: the TPAMI paper, the demo, the CNNs and the 2017 challenge; its download page holds, as of October 2026, a request form for the legacy Places365 and Places205 data.
• Places205 page: the 2014 release, its CNNs and the indoor/outdoor labels of the 205 categories.
• Places365-CNNs: pretrained models, the category list, indoor/outdoor labels, the adjective list and training scripts.
• Places365 development kit: split statistics, submission format and the evaluation routine.
• Places205 leaderboard (archived copy of September 2022; the live page returned a server error in October 2026), and the 2015 and 2016 challenge results.