SPICE: Semantic Propositional Image Caption Evaluation

Peter AndersonBasura FernandoMark JohnsonStephen Gould

article2016ECCV2,374 citations

Proposes SPICE, a scene-graph-based image caption evaluation metric that correlates significantly better with human judgment than traditional n-gram methods while enabling targeted diagnostic evaluation of semantic capabilities like counting and color comprehension.

Listen

Automated image captioning has advanced rapidly, yet evaluating generated captions remains a critical bottleneck. Standard automated metrics such as Bleu, ROUGE, CIDEr, and METEOR rely heavily on surface-level n-gram overlap. Because matching phrases is neither necessary nor sufficient to convey meaning, these conventional metrics frequently misjudge caption quality, penalize descriptive detail, and fail to reward captions that humans prefer.

The article introduces and evaluates SPICE (Semantic Propositional Image Caption Evaluation), a novel automated evaluation metric designed to measure caption quality based on semantic propositional content rather than phrase overlap. The authors set out to demonstrate that evaluating the semantic content of captionsspecifically the objects, attributes, and relationships they describebetter correlates with human judgment than existing n-gram metrics.

To achieve this, the approach converts both candidate and human reference captions into semantic scene graphs using dependency parsing and linguistic transformation rules. These scene graphs represent sentences as structured logical tuples denoting objects, their attributes, and their relationships. The metric then computes an F-score (a balanced measure of precision and recall) over matching tuples between the candidate and reference graphs, utilizing synonym matching to recognize related terms. The authors tested SPICE against standard metrics across four established benchmark datasets, including evaluation data from the 2015 MS COCO Captioning Challenge covering 15 competition entries, human captions, and 255,000 human quality judgments.

The analysis yielded several key findings. First, SPICE achieved a system-level correlation of 0.88 with human quality judgments on the MS COCO dataset, dramatically outperforming METEOR (0.53) and CIDEr (0.43). Second, SPICE was the only automated metric that correctly ranked human-written descriptions above all machine-generated entries and rewarded appropriate descriptive detail. Third, on pairwise classification tasks between competing models (PASCAL-50S dataset), SPICE showed the highest accuracy (68.2%) in identifying which model humans preferred. Finally, because the metric decomposes into explicit semantic categories, it enables detailed diagnostic evaluation; for example, subcategory analysis revealed that while top models effectively identified colors (sometimes surpassing human baselines), none demonstrated a meaningful capability to count objects accurately.

These findings indicate that evaluating semantic meaning aligns automated benchmarks much closer to genuine human perception. For research and technical leaders, adopting semantic evaluation mitigates the risk of selecting models that merely mimic superficial phrasing without understanding scene composition. It also provides actionable diagnostic insights into specific model weaknesses, such as counting or relational reasoning, which traditional metrics obscure.

Organizations developing or deploying vision-language systems should integrate SPICE alongside existing tools to benchmark model performance and guide targeted system improvements. Where caption fluency and grammatical correctness are also critical risks, practitioners should combine SPICE with fluency measures (such as surprisal metrics) in an ensemble evaluation. Future work should focus on integrating more advanced semantic parsers to further enhance metric precision and establishing upper-bound performance benchmarks against expert human semantic parsing.

Confidence in these findings is high for system-level ranking and model comparison on natural image datasets. However, stakeholders should note two primary limitations: SPICE relies on automated dependency parsing, which can introduce errors, and it does not explicitly penalize grammatical fluency issues on its own.

Cover for SPICE: Semantic Propositional Image Caption Evaluation

Abstract

There is considerable interest in the task of automatically generating image captions. However, evaluation is challenging. Existing automatic evaluation metrics are primarily sensitive to n-gram overlap, which is neither necessary nor sufficient for the task of simulating human judgment. We hypothesize that semantic propositional content is an important component of human caption evaluation, and propose a new automated caption evaluation metric defined over scene graphs coined SPICE. Extensive evaluations across a range of models and datasets indicate that SPICE captures human judgments over model-generated captions better than other automatic metrics (e.g., system-level correlation of 0.88 with human judgments on the MS COCO dataset, versus 0.43 for CIDEr and 0.53 for METEOR). Furthermore, SPICE can answer questions such as which caption-generator best understands colors?' and can caption-generators count?'

Table of Contents

  • 1 Introduction
  • 2 Background and Related Work
  • 2.1 Caption Evaluation Metrics
  • 2.2 Semantic Graphs
  • 3 SPICE Metric
  • 3.1 Semantic Parsing—Captions to Scene Graphs
  • 3.2 F-score Calculation
  • 3.3 Gameability
  • 4 Experiments
  • 4.1 Datasets
  • Microsoft COCO 2014.
  • Flickr 8K.
  • Composite Dataset.
  • PASCAL-50S
  • 4.2 System-Level Correlation
  • 4.3 Color Perception, Counting and Other Questions
  • 4.4 Caption-Level Correlation
  • 5 Conclusion and Future Work
  • References

Knowls

  1. Knowl 1 — SPICE Metric and Semantic Proposition F-Score

    model/method

    The Semantic Propositional Image Caption Evaluation (SPICE) metric measures caption quality by comparing semantic propositions extracted from candidate and reference captions represented as scene graphs. Given a candidate caption cc and a set of reference captions S={s1,,sm}S = \{s_1, \dots, s_m\}, SPICE constructs candidate scene graph G(c)G(c) and a combined reference scene graph G(S)=siSG(si)G(S) = \bigcup_{s_i \in S} G(s_i) where synonymous object nodes are merged.

    A function TT maps a scene graph G=O,E,KG = \langle O, E, K \rangle to a set of logical tuples:

    T(G)OEKT(G) \triangleq O \cup E \cup K

    where OO are 1-tuples representing object mentions (e.g., (girl)(girl)), KK are 2-tuples representing object attributes (e.g., (girl,young)(girl, young)), and EE are 3-tuples representing binary relations between objects (e.g., (girl,on-top-of,court)(girl, \text{on-top-of}, court)).

    Let \otimes denote the binary matching operator between two sets of tuples, where two tuples match if their lemmatized word forms match or belong to the same WordNet synset. No partial credit is given if only one element of a multi-element tuple matches. Precision P(c,S)P(c, S), recall R(c,S)R(c, S), and SPICE are defined as:

    P(c,S)=T(G(c))T(G(S))T(G(c))P(c, S) = \frac{|T(G(c)) \otimes T(G(S))|}{|T(G(c))|}

    R(c,S)=T(G(c))T(G(S))T(G(S))R(c, S) = \frac{|T(G(c)) \otimes T(G(S))|}{|T(G(S))|}

    SPICE(c,S)=F1(c,S)=2P(c,S)R(c,S)P(c,S)+R(c,S)\text{SPICE}(c, S) = F_1(c, S) = \frac{2 \cdot P(c, S) \cdot R(c, S)}{P(c, S) + R(c, S)}

    SPICE is bounded in [0,1][0, 1] and relies strictly on intra-sentence semantic tuples without requiring corpus-level statistics (such as tf-idf).

  2. Knowl 2 — Scene Graph Representation for Image Captions

    definition

    For an image caption cc, its semantic content is formalized as an open-world semantic scene graph:

    G(c)=O(c),E(c),K(c)G(c) = \langle O(c), E(c), K(c) \rangle

    where:

    • O(c)CO(c) \subseteq C is the set of object mentions in cc, drawn from an open-world set of object classes CC (which includes both discrete objects and stuff categories such as grass or sky);
    • E(c)O(c)×R×O(c)E(c) \subseteq O(c) \times R \times O(c) is the set of directed edges representing relations between object instances, drawn from relation types RR;
    • K(c)O(c)×AK(c) \subseteq O(c) \times A is the set of attributes associated with object instances, drawn from attribute types AA.

    Unlike visual grounding scene graphs, multiple instances of an object class are not duplicated as distinct nodes; instead, numerical quantities and counts are represented as attributes in K(c)K(c) associated with the corresponding object node.

  3. Knowl 3 — Two-Stage Semantic Parsing Pipeline for SPICE

    model/method

    To parse an image caption into a semantic scene graph G(c)G(c), SPICE uses a two-stage parsing pipeline:

    1. Syntactic Dependency Parsing: A pre-trained Probabilistic Context-Free Grammar (PCFG) dependency parser converts the raw text caption into a dependency syntax tree using Universal Dependency relations.
    2. Dependency Tree to Scene Graph Mapping: A rule-based parser transforms the dependency tree into a scene graph through three post-processing steps (simplifying quantificational modifiers, resolving pronouns, and handling plural nouns) followed by nine linguistic extraction rules to yield lemmatized objects, attributes, and relations.

    SPICE adapts standard textual scene graph parsers with two specific modifications tailored for caption evaluation:

    • Plural noun node duplication is eliminated; numeric modifiers (such as counts) are encoded directly as attributes of the object node so that an incorrect numeric modifier counts as only a single error.
    • An additional linguistic rule ensures that all extracted nouns appear as object nodes in the scene graph even if no relations or attributes are attached to them.
  4. Knowl 4 — System-Level Correlation with Human Judgments on MS COCO

    data/table

    On the 2015 MS COCO Captioning Challenge evaluation consisting of 15 automated systems plus human and random baselines (evaluated on a 1000-image subset with 40 reference captions each), SPICE demonstrates substantially higher Pearson correlation ρ\rho with human judgments than standard n-gram metrics across multiple evaluation dimensions:

    Metric M1 M2 M3 M4 M5
    ρ\rho p-val ρ\rho p-val ρ\rho p-val ρ\rho p-val ρ\rho p-val
    Bleu-1 0.24 (0.369) 0.29 (0.271) 0.72 (0.002) -0.54 (0.030) 0.44 (0.091)
    Bleu-4 0.05 (0.862) 0.10 (0.703) 0.58 (0.018) -0.63 (0.010) 0.30 (0.265)
    ROUGE-L 0.15 (0.590) 0.20 (0.469) 0.65 (0.006) -0.55 (0.030) 0.38 (0.142)
    METEOR 0.53 (0.036) 0.57 (0.022) 0.86 (0.000) -0.10 (0.710) 0.74 (0.001)
    CIDEr 0.43 (0.097) 0.47 (0.070) 0.81 (0.000) -0.21 (0.430) 0.65 (0.007)
    SPICE-exact 0.84 (0.000) 0.86 (0.000) 0.90 (0.000) 0.39 (0.000) 0.95 (0.000)
    SPICE 0.88 (0.000) 0.89 (0.000) 0.89 (0.000) 0.46 (0.070) 0.97 (0.000)

    Dimensions evaluated:

    • M1: Percentage of captions evaluated as better or equal to human captions.
    • M2: Percentage of captions that pass the Turing Test.
    • M3: Average caption correctness on a scale of 1–5.
    • M4: Average caption detail on a scale of 1–5.
    • M5: Percentage of captions similar to human descriptions.

    SPICE is the only metric that positively correlates with caption detailedness (M4, ρ=0.46\rho = 0.46) whereas n-gram metrics negatively correlate with detail. SPICE is also the only metric that ranks human-generated captions first among all submissions (CIDEr ranks human captions 7th; METEOR ranks them 4th).

  5. Knowl 5 — Diagnostic Semantic Subcategory Decomposition in SPICE

    model/method

    Because SPICE operates over discrete sets of logical propositions T(G)=OEKT(G) = O \cup E \cup K, precision, recall, and F-scores can be broken down into semantic subcategories:

    • Objects: Evaluated over 1-tuples O(c)O(c).
    • Relations: Evaluated over 3-tuples E(c)E(c).
    • Attributes: Evaluated over 2-tuples K(c)K(c).
    • Attribute Subcategories: Using predefined word lists, attribute tuples are partitioned into specialized subcategories:
      • Color: Attribute tuples containing color terms.
      • Count: Attribute tuples containing cardinal number words from one to ten.
      • Size: Attribute tuples containing size-related adjectives.

    This decomposition allows fine-grained diagnostic evaluation to isolate specific strengths and error modes across different caption generation models.

  6. Knowl 6 — Model Diagnostic Performance on Subcategories (Color, Count, and Size)

    empirical result

    Subcategory F-score evaluation of entries from the 2015 MS COCO Captioning Challenge reveals distinct model competencies:

    • Color Perception: Leading models trained with dedicated visual concept detectors (e.g., MSR at 0.063, Google at 0.060, MSR Captivator at 0.054) match or outperform the human reference baseline (0.055) in generating correct color attribute tuples.
    • Counting Ability: Automated caption generation models universally fail to count accurately. While the human baseline achieves an F-score of 0.095 on count tuples, automated models achieve scores near zero (the top model achieves 0.033, while the majority achieve 0.008\le 0.008), indicating that contemporary models lack object enumeration capabilities.
    • Size Attributes: Human captions achieve an F-score of 0.026 on size tuples, whereas the highest-performing model achieves 0.019 (MSR) and most models achieve 0.010\le 0.010.
  7. Knowl 7 — Caption-Level Correlation on Flickr 8K and Composite Datasets

    empirical result

    When evaluated at the individual caption level using Kendall's rank correlation coefficient τ\tau against human quality scores:

    • On the Flickr 8K dataset (5,822 captions evaluated by native speakers on a 1–4 scale), SPICE achieves τ=0.45\tau = 0.45, outperforming CIDEr (0.44), METEOR (0.42), ROUGE-L (0.32), Bleu-1 (0.32), and Bleu-4 (0.14). The inter-human rank correlation on this dataset is 0.73.
    • On the Composite dataset (11,985 captions from Flickr 8K, Flickr 30K, and MS COCO scored on a 1–5 scale), SPICE achieves τ=0.39\tau = 0.39, outperforming CIDEr (0.36), METEOR (0.35), ROUGE-L (0.28), Bleu-1 (0.26), and Bleu-4 (0.18).

    All reported caption-level correlation differences have p<0.001p < 0.001.

  8. Knowl 8 — Pairwise Caption Classification Accuracy on PASCAL-50S

    empirical result

    On the PASCAL-50S dataset (1,000 images with 50 captions each, tested across 4,000 candidate sentence pairs), evaluation metrics are assessed by classification accuracy in matching human preference between candidate sentences BB and CC against reference sentences. Candidate pairs are categorized into Human-Correct (HC), Human-Incorrect (HI), Human-Model (HM), and Model-Model (MM).

    Using 5 reference captions:

    • On Model-Model (MM) pairs (the core task of distinguishing between two automated model captions), SPICE achieves the highest accuracy at 68.2%, outperforming METEOR (66.8%), CIDEr (64.6%), ROUGE-L (60.3%), Bleu-1 (60.1%), and Bleu-2 (58.0%).
    • Across all pairs (All), METEOR achieves 80.8%, CIDEr achieves 78.9%, and SPICE achieves 78.8%.
    • METEOR performs highest on HM (94.2%) and HI (98.1%) pairs, while Bleu-1 performs highest on HC pairs (64.9%).
  9. Knowl 9 — Fluency Invariance and Grammaticality Assumption in SPICE

    limitation

    SPICE evaluates only the semantic propositional content (objects, attributes, relations) present in image captions and is completely invariant to grammatical fluency, syntax, and natural word ordering. Consequently, SPICE implicitly assumes candidate captions are well-formed natural language sentences and could be vulnerable to gaming by models generating unstructured bags of valid semantic propositions unless combined with a complementary fluency metric (e.g., language model surprisal).

Coverage note — None was omitted; all primary methodological components, formal mathematical formulations, empirical evaluations (COCO system-level, Flickr 8K/Composite caption-level, PASCAL-50S pairwise classification), diagnostic decomposition findings, and stated limitations are fully represented.

References

  1. 1.Donahue, J., Hendricks, L.A., Guadarrama, S., Rohrbach, M., Venugopalan, S., Saenko, K., Darrell, T.: Long-term recurrent convolutional networks for visual recognition and description. In: CVPR (2015)
  2. 2.Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A.C., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention (2015). arXiv preprint arXiv:1502.03044
  3. 3.Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking task: data, models and evaluation metrics. JAIR 47, 853–899 (2013)
  4. 4.Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: new similarity metrics for semantic inference over event descriptions. TACL 2, 67–78 (2014)
  5. 5.Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014, Part V. LNCS, vol. 8693, pp. 740–755. Springer, Heidelberg (2014)
  6. 6.Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft COCO captions: Data collection and evaluation server (2015). arXiv preprint arXiv:1504.00325
  7. 7.Kulkarni, G., Premraj, V., Ordonez, V., Dhar, S., Li, S., Choi, Y., Berg, A.C., Berg, T.L.: Babytalk: understanding and generating simple image descriptions. PAMI 35(12), 2891–2903 (2013)
  8. 8.Elliott, D., Keller, F.: Comparing automatic evaluation measures for image description. In: ACL, pp. 452–457 (2014)
  9. 9.Bernardi, R., Cakici, R., Elliott, D., Erdem, A., Erdem, E., Ikizler-Cinbis, N., Keller, F., Muscat, A., Plank, B.: Automatic description generation from images: a survey of models, datasets, and evaluation measures. JAIR 55, 409–442 (2016)
  10. 10.Papineni, K., Roukos, S., Ward, T., Zhu, W.: Bleu: a method for automatic evaluation of machine translation. In: ACL (2002)
  11. 11.Lin, C.Y.: ROUGE: a package for automatic evaluation of summaries. In: ACL Workshop, pp. 25–26 (2004)
  12. 12.Vedantam, R., Zitnick, C.L., Parikh, D.: CIDEr: consensus-based image description evaluation. In: CVPR (2015)
  13. 13.Denkowski, M., Lavie, A.: Meteor universal: language specific translation evaluation for any target language. In: EACL 2014 Workshop on Statistical Machine Translation (2014)
  14. 14.Giménez, J., Màrquez, L.: Linguistic features for automatic evaluation of heterogenous MT systems. In: ACL Second Workshop on Statistical Machine Translation
  15. 15.Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D.A., Bernstein, M.S., Fei-Fei, L.: Image retrieval using scene graphs. In: CVPR (2015)
  16. 16.Schuster, S., Krishna, R., Chang, A., Fei-Fei, L., Manning, C.D.: Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In: EMNLP 4th Workshop on Vision and Language (2015)
  17. 17.Wang, C., Xue, N., Pradhan, S.: A transition-based algorithm for AMR parsing. In: HLT-NAACL (2015)
  18. 18.Lin, D., Fidler, S., Kong, C., Urtasun, R.: Visual semantic search: retrieving videos via complex textual queries. In: CVPR (2014)
  19. 19.Klein, D., Manning, C.D.: Accurate unlexicalized parsing. In: ACL (2003)
  20. 20.De Marneffe, M.C., Dozat, T., Silveira, N., Haverinen, K., Ginter, F., Nivre, J., Manning, C.D.: Universal stanford dependencies: a cross-linguistic typology. LREC 14, 4585–4592 (2014)
  21. 21.Lo, C.k., Tumuluru, A.K., Wu, D.: Fully automatic semantic MT evaluation. In: ACL Seventh Workshop on Statistical Machine Translation (2012)
  22. 22.Pradhan, S.S., Ward, W., Hacioglu, K., Martin, J.H., Jurafsky, D.: Shallow semantic parsing using support vector machines. In: HLT-NAACL, pp. 233–240 (2004)
  23. 23.Ellebracht, L., Ramisa, A., Swaroop, P., Cordero, J., Moreno-Noguer, F., Quattoni, A.: Semantic tuples for evaluation of image sentence generation. In: EMNLP 4th Workshop on Vision and Language (2015)
  24. 24.Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitt, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation (AMR) 1.0 specification. In: EMNLP, pp. 1533–1544 (2012)
  25. 25.Flanigan, J., Thomson, S., Carbonell, J., Dyer, C., Smith, N.A.: A discriminative graph-based parser for the abstract meaning representation. In: ACL (2014)
  26. 26.Werling, K., Angeli, G., Manning, C.: Robust subgraph generation improves abstract meaning representation parsing. In: ACL (2015)
  27. 27.Cai, S., Knight, K.: Smatch: an evaluation metric for semantic feature structures. In: ACL (2), pp. 748–752 (2013)
  28. 28.Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: collecting region-to-phrase correspondences for richer image-to-sentence models. In: CVPR, pp. 2641–2649 (2015)
  29. 29.Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., Bernstein, M., Fei-Fei, L.: Visual genome: Connecting language and vision using crowdsourced dense image annotations (2016). arXiv preprint arXiv:1602.07332
  30. 30.Torralba, A., Efros, A.A.: Unbiased look at dataset bias. In: CVPR, June 2011
  31. 31.Hale, J.: A probabilistic earley parser as a psycholinguistic model. In: NAACL, pp. 1–8 (2001)
  32. 32.Levy, R.: Expectation-based syntactic comprehension. Cognition 106(3), 1126–1177 (2008)
  33. 33.Stanojević, M., Kamran, A., Koehn, P., Bojar, O.: Results of the WMT15 metrics shared task. In: ACL Tenth Workshop on Statistical Machine Translation, pp. 256–273 (2015)
  34. 34.Machacek, M., Bojar, O.: Results of the WMT14 metrics shared task. In: ACL Ninth Workshop on Statistical Machine Translation, pp. 293–301 (2014)
  35. 35.Aditya, S., Yang, Y., Baral, C., Fermuller, C., Aloimonos, Y.: From images to sentences through scene description graphs using commonsense reasoning and knowledge (2015). arXiv preprint arXiv:1511.03292
  36. 36.Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)
  37. 37.Rashtchian, C., Young, P., Hodosh, M., Hockenmaier, J.: Collecting image annotations using Amazon’s Mechanical Turk. In: HLT-NAACL, pp. 139–147 (2010)
  38. 38.Fang, H., Gupta, S., Iandola, F.N., Srivastava, R., Deng, L., Dollar, P., Gao, J., He, X., Mitchell, M., Platt, J.C., Zitnick, C.L., Zweig, G.: From captions to visual concepts and back. In: CVPR (2015)
  39. 39.Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and tell: a neural image caption generator. In: CVPR (2015)
  40. 40.Devlin, J., Cheng, H., Fang, H., Gupta, S., Deng, L., He, X., Zweig, G., Mitchell, M.: Language models for image captioning: The quirks and what works (2015). arXiv preprint arXiv:1505.01809
  41. 41.Mao, J., Wei, X., Yang, Y., Wang, J., Huang, Z., Yuille, A.L.: Learning like a child: fast novel visual concept learning from sentence descriptions of images. In: CVPR, pp. 2533–2541 (2015)
  42. 42.Devlin, J., Gupta, S., Girshick, R.B., Mitchell, M., Zitnick, C.L.: Exploring nearest neighbor approaches for image captioning (2015). arXiv preprint arXiv:1505.04467
  43. 43.Mao, J., Xu, W., Yang, Y., Wang, J., Huang, Z., Yuille, A.: Deep captioning with multimodal recurrent neural networks (m-rnn) (2014). arXiv preprint arXiv:1412.6632
  44. 44.Kolár, M., Hradis, M., Zemcík, P.: Technical report: Image captioning with semantically similar images (2015). arXiv preprint arXiv:1506.03995
  45. 45.Kiros, R., Salakhutdinov, R., Zemel, R.S.: Multimodal neural language models. ICML 14, 595–603 (2014)

Citation

MLA
Anderson, P., et al. “SPICE: Semantic Propositional Image Caption Evaluation”. Lecture Notes in Computer Science, Springer International Publishing, 2016, pp. 382–98, https://doi.org/10.1007/978-3-319-46454-1_24.
APA
Anderson, P., Fernando, B., Johnson, M., & Gould, S. (2016). SPICE: Semantic Propositional Image Caption Evaluation. In Lecture Notes in Computer Science (pp. 382–398). Springer International Publishing. https://doi.org/10.1007/978-3-319-46454-1_24
Chicago
Anderson, P., B. Fernando, M. Johnson, and S. Gould. 2016. “SPICE: Semantic Propositional Image Caption Evaluation”. In Lecture Notes in Computer Science. Springer International Publishing. https://doi.org/10.1007/978-3-319-46454-1_24.
Harvard
Anderson, P. et al. (2016) “SPICE: Semantic Propositional Image Caption Evaluation”, Lecture Notes in Computer Science. Springer International Publishing, pp. 382–398. Available at: https://doi.org/10.1007/978-3-319-46454-1_24.
Vancouver
1. Anderson P, Fernando B, Johnson M, Gould S (2016) SPICE: Semantic Propositional Image Caption Evaluation. In: Lecture Notes in Computer Science. Springer International Publishing, pp 382–398

BibTeX

@inbook{Anderson_2016, title={SPICE: Semantic Propositional Image Caption Evaluation}, ISBN={9783319464541}, ISSN={1611-3349}, url={http://dx.doi.org/10.1007/978-3-319-46454-1_24}, DOI={10.1007/978-3-319-46454-1_24}, booktitle={Computer Vision – ECCV 2016}, publisher={Springer International Publishing}, author={Anderson, Peter and Fernando, Basura and Johnson, Mark and Gould, Stephen}, year={2016}, pages={382–398} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF